
I Create • I Code • I Build I • Win. Contributor @SeismicSys
Earth
Joined December 2024
- Tweets3K
- Following603
- Followers1.2K
- Likes20.8K
Pizzy Lee retweeted
Every month, Senegalese workers in Paris, Madrid, and New York send money home to Dakar for pharmacy prescriptions, school fees, or bags of cement.
Standard remittance rails charge steep fees, arrive as fungible cash that often gets diverted to other bills, and leave families sending blurry WhatsApp photos as proof of payment.
Meet KeurBon ("keur" = household in Wolof, "bon" = voucher in French): non-custodial, purpose-locked vouchers built on Bitcoin Cashu and Nostr.
Zero custodial servers. Zero personal data stored or broadcast.
Here is how a voucher moves from Paris to a neighborhood counter in Dakar:
1. Vendor-locked issuance (Cashu NUT-11 P2PK)
The sender enters an amount in FCFA, picks a Dakar merchant (such as Pharmacie Medina Rue 6), and sets an expiry window. KeurBon mints real Cashu proofs locked strictly to that merchant's secp256k1 public key, with automatic NUT-02 swap fee headroom and the sender's refund key embedded.
2. 100% offline counter inspection (Cashu NUT-12 DLEQ)
Before touching flaky 2G/3G mobile data, the Dakar merchant verifies the NUT-11 pubkey lock and validates every NUT-12 DLEQ mint blind signature locally against the cached keyset.
3. Enforced live at the mint
If an unauthorized third party tries to divert or cash out the voucher, submitting a forged key fails directly at the mint with HTTP 400 (code 20008: P2PK spend conditions are not met).
4. Schnorr claim + zero-PII Nostr receipts (kind 30078)
The authorized merchant signs with their key, claims the sats, and publishes a Nostr kind 30078 receipt to nos.lol and relay.primal.net. Anyone can verify the BIP-340 signature independently. The receipt contains only the voucher SHA-256 hash, vendor npub, category, and amount. Zero family names, phone numbers, or addresses.
5. Automatic expiry safety net (NUT-07 + NUT-11 refund)
If a voucher goes unused past its locktime, the sender audits live UNSPENT state in the vault and reclaims the satoshis in one click. No support tickets, no middlemen.
Built for real Dakar conditions:
- 5 real routes in a high-contrast daylight paper UI
- French-first copy with an instant English toggle
- Self-hosted fonts (zero external font CDN blocks on slow mobile data)
- 16/16 passing unit and live Testnut cryptographic tests
- One-command Docker setup, MIT open source
Built for the Dakar Bitcoin Hack by BTrust 2026.
Try it live: keurbon.vercel.app/
Source code (MIT): github.com/Pizzylee001/KeurB…
Every month, Senegalese workers in Paris, Madrid, and New York send money home to Dakar for pharmacy prescriptions, school fees, or bags of cement.
Standard remittance rails charge steep fees, arrive as fungible cash that often gets diverted to other bills, and leave families sending blurry WhatsApp photos as proof of payment.
Meet KeurBon ("keur" = household in Wolof, "bon" = voucher in French): non-custodial, purpose-locked vouchers built on Bitcoin Cashu and Nostr.
Zero custodial servers. Zero personal data stored or broadcast.
Here is how a voucher moves from Paris to a neighborhood counter in Dakar:
1. Vendor-locked issuance (Cashu NUT-11 P2PK)
The sender enters an amount in FCFA, picks a Dakar merchant (such as Pharmacie Medina Rue 6), and sets an expiry window. KeurBon mints real Cashu proofs locked strictly to that merchant's secp256k1 public key, with automatic NUT-02 swap fee headroom and the sender's refund key embedded.
2. 100% offline counter inspection (Cashu NUT-12 DLEQ)
Before touching flaky 2G/3G mobile data, the Dakar merchant verifies the NUT-11 pubkey lock and validates every NUT-12 DLEQ mint blind signature locally against the cached keyset.
3. Enforced live at the mint
If an unauthorized third party tries to divert or cash out the voucher, submitting a forged key fails directly at the mint with HTTP 400 (code 20008: P2PK spend conditions are not met).
4. Schnorr claim + zero-PII Nostr receipts (kind 30078)
The authorized merchant signs with their key, claims the sats, and publishes a Nostr kind 30078 receipt to nos.lol and relay.primal.net. Anyone can verify the BIP-340 signature independently. The receipt contains only the voucher SHA-256 hash, vendor npub, category, and amount. Zero family names, phone numbers, or addresses.
5. Automatic expiry safety net (NUT-07 + NUT-11 refund)
If a voucher goes unused past its locktime, the sender audits live UNSPENT state in the vault and reclaims the satoshis in one click. No support tickets, no middlemen.
Built for real Dakar conditions:
- 5 real routes in a high-contrast daylight paper UI
- French-first copy with an instant English toggle
- Self-hosted fonts (zero external font CDN blocks on slow mobile data)
- 16/16 passing unit and live Testnut cryptographic tests
- One-command Docker setup, MIT open source
Built for the Dakar Bitcoin Hack by BTrust 2026.
Try it live: keurbon.vercel.app/
Source code (MIT): github.com/Pizzylee001/KeurB…
Three days ago we asked you to build search that doesn't freeze the day it ships.
Hundreds of you did.
Thank you to everyone who built something. Winners soon.
Pizzy Lee retweeted
Those who leave everyone in Gods hand will soon see Gods hand in everything 🤲🏽
Happy international men’s day, y’all goated 🐐
Pizzy Lee retweeted
if you load a standard generative 3d cabinet into pybullet, mujoco, or isaac sim, a robot arm cannot open the drawer. the handle, drawer front, and cabinet body are fused into a single static triangle mesh.
aholo lux3d solves the hardest geometry bottleneck with its g1-turbo + part split pipeline, separating a single mesh into clean, independent sub-meshes.
yet between a part-split glb and a working robotics simulation sits a second gap: a separated mesh still has no joint axes, no travel limits, no grasp frame for an end-effector, and no check on whether a robot arm will collide with the object halfway through opening it.
we built jointbench to close that gap.
what jointbench does:
1. lux3d pipeline foundry (/foundry)
generate raw 3d assets via aholo lux3d g1-turbo, trigger single-slot part split jobs, or drop in a local studio glb. our 3d curtain comparator slices the raw monolithic mesh directly against the exploded lux3d part-split links so you can inspect triangle counts, bounding boxes, and vertex normals before rigging.
2. 3d kinematic calibration bench (/bench/[id])
bind each lux3d sub-mesh to a base or moving role and configure either a prismatic slide (+x, +y, +z in meters) or a revolute hinge (+x, +y, +z in degrees) with an exact 3d pivot origin.
3. 3-dof robotic arm sweep & one-click collision repair
a built-in 3-dof planar manipulator (0.85 m max reach) tracks the moving handle across 24 trajectory waypoints in real time:
- bedside drawer cabinet (lux3d raw #8414321 -> part split #8419896, 3 links, 55,048 tris): prismatic +z slide from 0.00 m to 0.24 m passes at 61% arm reach with +80 mm body clearance.
- retro countertop microwave (lux3d raw #8419901 -> part split #8420171, 4 links, 48,268 tris): revolute -y door swing from 0 deg to 88 deg passes at 61% arm reach with +45 mm clearance.
- brass-clasped storage chest (lux3d raw #8419905 -> part split #8420433, 6 links, 52,776 tris): revolute -x lid swing to 125 deg flags a -59 mm rear collision at 99% arm reach. one click on "apply recommended safe limit" clamps the upper stop to 92 deg and restores a verified pass with +34 mm clearance.
4. simulator-ready urdf 1.0 zip export
every calibration updates a live urdf 1.0 xml tree with inertial tensors, visual and collision nodes, calibrated joint limits, and a dedicated fixed grasp_handle_link frame for end-effector planners. one click downloads a self-contained zip bundling the urdf, split glb, calibration-report.json, and ready-to-run loader scripts for pybullet, mujoco, and nvidia isaac sim.
built with next.js, three.js, and the aholo lux3d api for the manycore aholo global 3d ai hackathon (embodied ai application track).
live app: jointbench.vercel.app
code: github.com/Pizzylee001/joint…
#Manycorehackathon @ManycoreTech
Pizzy Lee retweeted
gm gm chat
tracemend made it onto @openservai's featured builds.
built it because eval gates block bad agent outputs and stop there. tracemend pulls live Reppo evidence, mends the broken claims with serv, verifies the score lift, and exports clean dpo training pairs.
try it today!
Most AI evaluation tools reject a bad agent output, hand you a score, and stop there. You still have to track down sources, rewrite broken claims, re-verify the lift, and format training data yourself.
I built TraceMend
Here is how it works across three workspaces:
1. /datanets (Reppo Evidence Registry)
Browse 19 live Reppo DataNets, check vote ratios and epoch emissions, verify IPFS pod provenance, and pin real evidence straight into your workbench.
2. /mend (Mend Bench + SERV Reasoning)
Submit agent answers, plans, traces, or artifacts to live Reppo Evaluation Markets (eval.reppo.xyz) or run an instant SERV Pre-Screen. When a draft gets flagged or blocked, TraceMend routes the critique and pinned pods through SERV Reasoning (Gpt 6 Luna) with immutable BRAID prompt caching, serv_prompt_guard, serv_shadow_agent verification, and strict JSON schema output.
Then inspect every edit with the Kintsugi Optical Seam diff comparator and verify the score lift across Accuracy, Grounding, Completeness, and Safety.
3. /vault (DPO Preference Vault)
Every repaired run compiles into an OpenAI and TRL-compatible DPO training pair, pairing the rejected output with the citation-verified fix for one-click .jsonl and Markdown receipt export.
Built for the @openservai Hackathon Edition 02 (Track 03: Reppo + SERV Reasoning)
Try it live: tracemend.vercel.app
Source code: github.com/Pizzylee001/trace…
if you load a standard generative 3d cabinet into pybullet, mujoco, or isaac sim, a robot arm cannot open the drawer. the handle, drawer front, and cabinet body are fused into a single static triangle mesh.
aholo lux3d solves the hardest geometry bottleneck with its g1-turbo + part split pipeline, separating a single mesh into clean, independent sub-meshes.
yet between a part-split glb and a working robotics simulation sits a second gap: a separated mesh still has no joint axes, no travel limits, no grasp frame for an end-effector, and no check on whether a robot arm will collide with the object halfway through opening it.
we built jointbench to close that gap.
what jointbench does:
1. lux3d pipeline foundry (/foundry)
generate raw 3d assets via aholo lux3d g1-turbo, trigger single-slot part split jobs, or drop in a local studio glb. our 3d curtain comparator slices the raw monolithic mesh directly against the exploded lux3d part-split links so you can inspect triangle counts, bounding boxes, and vertex normals before rigging.
2. 3d kinematic calibration bench (/bench/[id])
bind each lux3d sub-mesh to a base or moving role and configure either a prismatic slide (+x, +y, +z in meters) or a revolute hinge (+x, +y, +z in degrees) with an exact 3d pivot origin.
3. 3-dof robotic arm sweep & one-click collision repair
a built-in 3-dof planar manipulator (0.85 m max reach) tracks the moving handle across 24 trajectory waypoints in real time:
- bedside drawer cabinet (lux3d raw #8414321 -> part split #8419896, 3 links, 55,048 tris): prismatic +z slide from 0.00 m to 0.24 m passes at 61% arm reach with +80 mm body clearance.
- retro countertop microwave (lux3d raw #8419901 -> part split #8420171, 4 links, 48,268 tris): revolute -y door swing from 0 deg to 88 deg passes at 61% arm reach with +45 mm clearance.
- brass-clasped storage chest (lux3d raw #8419905 -> part split #8420433, 6 links, 52,776 tris): revolute -x lid swing to 125 deg flags a -59 mm rear collision at 99% arm reach. one click on "apply recommended safe limit" clamps the upper stop to 92 deg and restores a verified pass with +34 mm clearance.
4. simulator-ready urdf 1.0 zip export
every calibration updates a live urdf 1.0 xml tree with inertial tensors, visual and collision nodes, calibrated joint limits, and a dedicated fixed grasp_handle_link frame for end-effector planners. one click downloads a self-contained zip bundling the urdf, split glb, calibration-report.json, and ready-to-run loader scripts for pybullet, mujoco, and nvidia isaac sim.
built with next.js, three.js, and the aholo lux3d api for the manycore aholo global 3d ai hackathon (embodied ai application track).
live app: jointbench.vercel.app
code: github.com/Pizzylee001/joint…
#Manycorehackathon @ManycoreTech
Can an AI spot a misleading property ad before someone loses their life savings? 🏠📉
Picture a young couple saving for years for a house deposit gets an email with promises of a "50–100% return" on their money.
But they have no way to verify it.
The UK ad regulator (ASA) eventually ruled that exact claim misleading but only after someone complained. Regulators can't read every ad.
So I built Porkies (UK slang for lies) for the @openservai SERV Hackathon ( @reppo track) to automate compliance checks with verifiable proof.
How it works:
1️⃣ Code Calculator:
Verifies the actual math in code.
2️⃣ SERV Reasoning:
Evaluates the claim against evidence + blocks prompt injections.
3️⃣ Reppo Panel:
Independently reviews the verdict so the AI isn't marking its own homework. If flagged, it goes back to SERV for a second look.
4️⃣ Tamper-Proof Logs:
Every step is hash-chained and verifiable directly in your browser.
Honest Benchmarks (Zero Cherry-Picking):
I tested Porkies against 24 real claims from 9 UK ASA rulings (19 misleading, 5 fair) using the regulator's rulings as the answer key:
✅ 21/24 total agreement with the regulator (87.5%)
✅ 19/19 misleading claims caught (100% recall)
⚠️ 3/5 fair claims flagged as misleading (it errs on the side of caution)
⚠️ In this run, the Reppo panel missed those 3 false positives and added 1 false flag. So the app features a dedicated "What this does not prove" section for full transparency.
Test it live:
Paste any claim and its evidence to watch SERV and Reppo audit it in real time.
🔗 Live App: porkies.ibnweb3lab.workers.d…💻 Code: github.com/ibnweb3/porkies
Built with @openservai SERV Reasoning & @reppo. #SERVHackathon
gm gm chat
tracemend made it onto @openservai's featured builds.
built it because eval gates block bad agent outputs and stop there. tracemend pulls live Reppo evidence, mends the broken claims with serv, verifies the score lift, and exports clean dpo training pairs.
try it today!
Most AI evaluation tools reject a bad agent output, hand you a score, and stop there. You still have to track down sources, rewrite broken claims, re-verify the lift, and format training data yourself.
I built TraceMend
Here is how it works across three workspaces:
1. /datanets (Reppo Evidence Registry)
Browse 19 live Reppo DataNets, check vote ratios and epoch emissions, verify IPFS pod provenance, and pin real evidence straight into your workbench.
2. /mend (Mend Bench + SERV Reasoning)
Submit agent answers, plans, traces, or artifacts to live Reppo Evaluation Markets (eval.reppo.xyz) or run an instant SERV Pre-Screen. When a draft gets flagged or blocked, TraceMend routes the critique and pinned pods through SERV Reasoning (Gpt 6 Luna) with immutable BRAID prompt caching, serv_prompt_guard, serv_shadow_agent verification, and strict JSON schema output.
Then inspect every edit with the Kintsugi Optical Seam diff comparator and verify the score lift across Accuracy, Grounding, Completeness, and Safety.
3. /vault (DPO Preference Vault)
Every repaired run compiles into an OpenAI and TRL-compatible DPO training pair, pairing the rejected output with the citation-verified fix for one-click .jsonl and Markdown receipt export.
Built for the @openservai Hackathon Edition 02 (Track 03: Reppo + SERV Reasoning)
Try it live: tracemend.vercel.app
Source code: github.com/Pizzylee001/trace…
built this for the @testmuai kane cli hackathon after getting burned by outdated quickstarts one too many times.
the hard part isn't knowing a setup guide broke. it's knowing why:
- did someone rename a button and forget to update the docs?
- or did the app itself break underneath them?
claimcheck figures that out from one real browser run, writes the readme fix, packages the bug for your coding agent, and re-checks until everything passes.
Most unit tests stay green even when your README is wrong. Someone renames "Generate API Token" to "Create Secret Key", or breaks an API route with an HTTP 500, and your onboarding docs silently fail the next user.
I built ClaimCheck (powered by @testmuai Kane CLI) to test product docs like code and repair what fails:
- Reads your README.md and turns each step into a real headless Chrome browser check
- Unpacks Kane CLI .evidence archives (screenshots, network logs, console errors) and inspects drifted UI elements with a 2x reticle
- Splits failures cleanly into two buckets:
1. Documentation Drift: writes the exact README.md PR diff so you can update the stale wording
2. Code Regression: packages the failing API route, HTTP 500 trace, and selector into a prompt any coding agent can fix
- Re-runs the check against the app until Exit 1 turns into Exit 0
Try the live app & built-in sandbox: claimcheck-nu.vercel.app
Code: github.com/Pizzylee001/Claim…
Built for the KaneCLI Online Hackathon.
Pizzy Lee retweeted
Most unit tests stay green even when your README is wrong. Someone renames "Generate API Token" to "Create Secret Key", or breaks an API route with an HTTP 500, and your onboarding docs silently fail the next user.
I built ClaimCheck (powered by @testmuai Kane CLI) to test product docs like code and repair what fails:
- Reads your README.md and turns each step into a real headless Chrome browser check
- Unpacks Kane CLI .evidence archives (screenshots, network logs, console errors) and inspects drifted UI elements with a 2x reticle
- Splits failures cleanly into two buckets:
1. Documentation Drift: writes the exact README.md PR diff so you can update the stale wording
2. Code Regression: packages the failing API route, HTTP 500 trace, and selector into a prompt any coding agent can fix
- Re-runs the check against the app until Exit 1 turns into Exit 0
Try the live app & built-in sandbox: claimcheck-nu.vercel.app
Code: github.com/Pizzylee001/Claim…
Built for the KaneCLI Online Hackathon.
Most unit tests stay green even when your README is wrong. Someone renames "Generate API Token" to "Create Secret Key", or breaks an API route with an HTTP 500, and your onboarding docs silently fail the next user.
I built ClaimCheck (powered by @testmuai Kane CLI) to test product docs like code and repair what fails:
- Reads your README.md and turns each step into a real headless Chrome browser check
- Unpacks Kane CLI .evidence archives (screenshots, network logs, console errors) and inspects drifted UI elements with a 2x reticle
- Splits failures cleanly into two buckets:
1. Documentation Drift: writes the exact README.md PR diff so you can update the stale wording
2. Code Regression: packages the failing API route, HTTP 500 trace, and selector into a prompt any coding agent can fix
- Re-runs the check against the app until Exit 1 turns into Exit 0
Try the live app & built-in sandbox: claimcheck-nu.vercel.app
Code: github.com/Pizzylee001/Claim…
Built for the KaneCLI Online Hackathon.
Pizzy Lee retweeted
Every trader knows the feeling of owning 4 "different" altcoins and still watching all 4 dump at the exact same speed when Bitcoin sneezes.
That is the exact blind spot ShockBook was built for.
Instead of giving you another green/red price chart, ShockBook puts a quantitative risk officer in front of your entire portfolio:
• Drag one slider to simulate a -15% flash crash, a liquidity crunch, or a perp funding squeeze
• Watch which "uncorrelated" bags quietly snap together toward r = 0.88 correlation
• Get the exact dollar trims into BTC and USDC that keep your drawdown under your limit before a wick hits
• Lock the plan into a verifiable audit ledger so you can track how much capital the rebalance actually saved
And for anyone building AI trading agents on @ryodigital, all of that risk logic runs behind the scenes as open portfolio tools your agent can check before ever entering a trade.
Test your own book live in 30 seconds:
shockbook.vercel.app
most crypto research tools score tokens in isolation.
you can hold 42% SOL and 28% AVAX, see bullish momentum on both, and have no idea their 168-hour correlation has climbed to 0.88. seventy percent of your portfolio is secretly sitting in the same trade. when BTC drops 15%, both positions collapse together and a $20,000 book takes an unhedged -$4,627 (-23.1%) hit, blowing straight past a $2,000 drawdown limit.
I built shockbook to solve that multi-token blind spot.
what shockbook does:
1. stress desk (/desk)
combines live RYO-CHAN market intelligence with 168-hour OKX hourly candles to compute your portfolio's pairwise correlation matrix, empirical BTC shock beta, and 48-hour shock confidence cone. when a shock breaches your risk cap, the minimum-variance solver computes exact rebalance trims into BTC and USDC, cutting that -$4,627 drawdown to -$1,836 and preserving +$2,791 of capital.
2. mandate audit ledger (/ledger)
seals every committed rebalance mandate into an immutable SQLite WAL ledger with a deterministic SHA-256 replay hash, tracking live counterfactual P&L against your unhedged starting weights.
3. composable MCP risk server (/skills)
extends RYO-CHAN's single-token tools with two multi-asset MCP skills (`portfolio_stress_test` and `correlation_guard`) over JSON-RPC 2.0 and REST, complete with a 60 RPM quota guard and explicit null safety so AI agents can verify portfolio-wide tail risk before sizing a trade.
why it matters:
traders and autonomous agents don't get wiped out by single-token thesis errors alone. they get wiped out when uncorrelated positions quietly converge to correlation 1.0 during a liquidity shock. shockbook turns isolated token signals into enforceable portfolio risk mandates.
built for the @ryodigital virtual hackathon.
try live at: shockbook.vercel.app
code: github.com/Pizzylee001/Shock…
Every trader knows the feeling of owning 4 "different" altcoins and still watching all 4 dump at the exact same speed when Bitcoin sneezes.
That is the exact blind spot ShockBook was built for.
Instead of giving you another green/red price chart, ShockBook puts a quantitative risk officer in front of your entire portfolio:
• Drag one slider to simulate a -15% flash crash, a liquidity crunch, or a perp funding squeeze
• Watch which "uncorrelated" bags quietly snap together toward r = 0.88 correlation
• Get the exact dollar trims into BTC and USDC that keep your drawdown under your limit before a wick hits
• Lock the plan into a verifiable audit ledger so you can track how much capital the rebalance actually saved
And for anyone building AI trading agents on @ryodigital, all of that risk logic runs behind the scenes as open portfolio tools your agent can check before ever entering a trade.
Test your own book live in 30 seconds:
shockbook.vercel.app
most crypto research tools score tokens in isolation.
you can hold 42% SOL and 28% AVAX, see bullish momentum on both, and have no idea their 168-hour correlation has climbed to 0.88. seventy percent of your portfolio is secretly sitting in the same trade. when BTC drops 15%, both positions collapse together and a $20,000 book takes an unhedged -$4,627 (-23.1%) hit, blowing straight past a $2,000 drawdown limit.
I built shockbook to solve that multi-token blind spot.
what shockbook does:
1. stress desk (/desk)
combines live RYO-CHAN market intelligence with 168-hour OKX hourly candles to compute your portfolio's pairwise correlation matrix, empirical BTC shock beta, and 48-hour shock confidence cone. when a shock breaches your risk cap, the minimum-variance solver computes exact rebalance trims into BTC and USDC, cutting that -$4,627 drawdown to -$1,836 and preserving +$2,791 of capital.
2. mandate audit ledger (/ledger)
seals every committed rebalance mandate into an immutable SQLite WAL ledger with a deterministic SHA-256 replay hash, tracking live counterfactual P&L against your unhedged starting weights.
3. composable MCP risk server (/skills)
extends RYO-CHAN's single-token tools with two multi-asset MCP skills (`portfolio_stress_test` and `correlation_guard`) over JSON-RPC 2.0 and REST, complete with a 60 RPM quota guard and explicit null safety so AI agents can verify portfolio-wide tail risk before sizing a trade.
why it matters:
traders and autonomous agents don't get wiped out by single-token thesis errors alone. they get wiped out when uncorrelated positions quietly converge to correlation 1.0 during a liquidity shock. shockbook turns isolated token signals into enforceable portfolio risk mandates.
built for the @ryodigital virtual hackathon.
try live at: shockbook.vercel.app
code: github.com/Pizzylee001/Shock…
Pizzy Lee retweeted
Just published a technical write-up on how I built ToolPath using open-source @hydra_db (`github.com/hydra-db/hydradb`) to cut MCP agent prompt context by 92.3% while surviving live tool outages.
Key engineering takeaways inside the post:
- Why flat 42-tool prompts (18,400 tokens, 52% accuracy) and static cosine top-K (64%, breaks on 504 timeouts) fail multi-step workflows
- Using CALL algo.SPpaths + CALL algo.degree + OpenCypher SET r.cost to route around failing endpoints live (v:103 -> v:104 in 1,380 tokens)
- How we handled local graph-node LAGraph/RUST_MIN_STACK wiring alongside serverless Vercel deployment (68% -> 96% accuracy over 20 steps)
Read the full breakdown.
Just published a technical write-up on how I built ToolPath using open-source @hydra_db (`github.com/hydra-db/hydradb`) to cut MCP agent prompt context by 92.3% while surviving live tool outages.
Key engineering takeaways inside the post:
- Why flat 42-tool prompts (18,400 tokens, 52% accuracy) and static cosine top-K (64%, breaks on 504 timeouts) fail multi-step workflows
- Using CALL algo.SPpaths + CALL algo.degree + OpenCypher SET r.cost to route around failing endpoints live (v:103 -> v:104 in 1,380 tokens)
- How we handled local graph-node LAGraph/RUST_MIN_STACK wiring alongside serverless Vercel deployment (68% -> 96% accuracy over 20 steps)
Read the full breakdown.








