Sahil Menon

Refusal can be stripped from an open-weight model by a quiet fine-tune. Whether that leaves a readable signature depends on where you look, and the obvious place, the prompt, reads chance.

Refusal in Llama-3.1-8B-Instruct is a single linear direction at layer 10: ablate it and refusal falls 92.5% → 0% with benign behaviour intact. A covert LoRA fine-tune strips it (98.8% → 0%), and a generation-time probe detects that tampering at AUROC 0.956 where a prompt-side probe reads chance. It holds at 1.000 against an adaptive attacker that jailbreaks the model 94% of the time. The same detector transfers unchanged to sleeper-agent backdoors (0.965) and emergent misalignment.

AI-assisted coding wastes tokens on bloated prompts with no feedback loop. PromptCoach turns that repeated waste into a one-time cost.

One shared analysis core behind a CLI and Chrome extension (TypeScript/Next.js + FastAPI, Drizzle ORM on Cloudflare D1). Coaching is opt-in via a review: prompt prefix, so ordinary prompts never leave the machine; a local bridge falls back to offline heuristics. Hosted analysis runs through Anthropic’s Batches API on Haiku at ~2% of a frontier model’s energy per task, with environmental impact reported as a sourced, uncertainty-labelled range. Local-first, no telemetry.

Keeping a market-making engine profitable as volatility regimes change, without overfitting to any one of them.

Built a three-tier market-making engine (take/clear/make) using Welford online mean, online AR(1) on price deviations, z-score tiered sizing, and asymmetric bid/ask anchoring. Built a separate trend-following MM with hardcoded-slope discovery, online OLS blending (70/30), and full-book order imbalance microprice adjustment.

Cut latency without changing a single matched trade, and prove the book still matches the naive version byte for byte.

A flat O(1) price-ladder array and intrusive free-list order pool (64-byte cache-aligned levels) replace std::map + std::list, cutting median latency 538→344 ns (~1.5×) and lifting throughput to 4.0 M ops/s. The rewrite stays byte-for-byte identical to the naive book and holds up against 2.8M replayed NASDAQ ITCH messages. Compiled to WebAssembly for a live in-browser demo.

A confident wrong DSCR costs a credit desk more than a flagged one. Most extraction benchmarks never test abstention, so they can’t tell the two apart.

Scores a credit-metric extractor on accuracy, grounding, and abstention. The model extracts line items only; ratios like leverage and DSCR come from plain Python, so every error traces to one source line. On 15 real UK charity filings, a Claude-backed extractor traced every answer to a verbatim quote and fabricated none, where a regex baseline answered nothing correctly. A CI gate fails the build when a metric regresses.

Kuhn poker has a closed-form solution, so a correct solver must reproduce its game value of −1/18. This solver checks itself against it.

CFR and CFR+ solvers built from the original papers drive exploitability to 9×10⁻⁴ / 1.5×10⁻³ chips/game on Kuhn and Leduc, recovering Kuhn’s closed-form value of −1/18 to within 4×10⁻⁷. CFR+ converges ~10× faster; the solved strategy wins +38 to +72 bb/100 vs fixed baselines. Playable in-browser, with JS/Python info-set-key parity self-checks.

A 10-billion-cell maze at full resolution would need terabytes of memory. This engine streams it in tens of MB and still holds a 3.3 ms p99 round trip with 100 concurrent players.

Eller’s algorithm with within-row strip parallelism and AVX2/BMI2 SIMD packing scales to ~92 Mcells/s on 8 cores in O(width) memory. FastAPI + Redis pub/sub fan-out across K8s pods (p99 3.3 ms at 100 concurrent bots); WebGL2 renderer batches an entire 64×64 chunk into one GPU draw call at 60 fps. One-command AWS deploy via Terraform + EKS.

Each game ships as a shareable Google Slides link, with no runtime, JavaScript, or server behind it.

1,000-state ceiling bounds exponential growth (Pac-Man scales as positions × 2ⁿ with n pellets). Token-bucket rate limiter at ≤50 API writes/min with 5 concurrent batch uploads generates ~500-state presentations in 1–3 min. Pygame-inspired 1920×1080 rendering API with 40+ colours and 3 themes; campaign system across 4 bundled games (491–600 states each).

YouTube lossy-compresses every uploaded video. Storing arbitrary binary data there without any corruption is the problem.

2×2 uniform pixel blocks survive ±127 DCT luma ringing in H.264/VP9. Three-tier Reed-Solomon ECC over GF(2⁸) (vectorised → Berlekamp-Massey → parallel) with byte interleaving for burst-error recovery. AES-256-GCM + PBKDF2-SHA256 encryption, zlib compression, and 38× faster encoding at 0.82–2.88 MB/s via NVENC/AMF/QSV hardware acceleration with YouTube OAuth2 upload.

Map full-colour images to text characters at 30 fps without losing perceptual detail.

O(1) nearest-colour lookup via a precomputed 32³ = 32,768-entry RGB quantisation table enables 30 fps video at ~2.1 ms/frame. Implements Floyd-Steinberg, Atkinson, and Bayer dithering, Sobel edge detection, and a particle drift system. Exports PNG, SVG, TXT, and WebM (via the MediaRecorder API) and reports live render time at 1920×1080.