Skip to content
View anamankev's full-sized avatar
🎯
Focusing
🎯
Focusing
0 followers · 1 following

Achievements

Achievement: Pull Shark

Achievements

Achievement: Pull Shark

Highlights

  • Pro

Block or report anamankev

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
anamankev/README.md
ascii-art (1)

Kelvin Anaman. ML systems built from the paper up, measured against the best libraries. Computer science and applied math, Brandeis class of 2030, Boston. Founder of Pulse Energies. Wien Scholar. A real HNSW index of 40 points built by hnsw-engine, drawn as three layers, with one search highlighted from the top layer down to the nearest neighbour.

The graph is real: an index of 40 points built by hnsw-engine, and one search through it, from the entry point down to the nearest neighbour.

I study computer science and applied mathematics at Brandeis (Wien Scholar, class of 2030) and I'm the founder of Pulse Energies, where we build software for electricity grids.

On GitHub I rebuild the systems I want to understand. I start from the paper, test hard, benchmark against the libraries everyone already trusts, and write down every gap, including the ones that don't flatter me.

LinkedIn · Email · Pulse Energies

Selected work

brier. A prediction-market exchange in Rust, built to a regulated venue's standard, with play money. 12.3 million orders matched per second on one core; 1,000 crashes under load with no acknowledged order lost; 10 million commands with no divergence from an independent reference engine. Where an order's time goes, log scale: matching p99 334 ns; matching with ledger checks p99 1.4 microseconds; durable disk flush median 3.95 ms; order to acknowledgment through the gateway p99 9.14 ms.

brier is a prediction-market exchange, named after the Brier score, built the way a regulated venue has to be and run with play money. It has a deterministic Rust matching engine and a double-entry ledger, held event for event to an independent Python reference engine. Its journal acknowledges no order before it is on disk, proven by killing it 1,000 times under load. It is live for the public on two free cloud servers, with about 100 open events across weather, science, economics, rates and sports. Markets settle from public data (NWS, USGS, BLS, FRED, NOAA and league results) with the evidence hashed into the journal, a market maker trades only through the public API, and surveillance detects wash trading, spoofing and insider-like trading. Every claim has a pre-registered test: the surveillance study reports two detectors that passed and two that did not. Live site · Measured results · Specification · Surveillance study

hnsw-engine. HNSW built from the original paper, twice. 4,447 queries per second at recall 0.95 or better on one million vectors; 39 times faster than exact search; 336 tests. Queries per second at recall 0.95 or better on SIFT-1M, one thread: exact search 114, hnsw-engine 4,447, hnswlib 6,193, FAISS HNSW 10,301.

hnsw-engine is HNSW approximate nearest-neighbour search, implemented from the original paper twice: a readable Python reference, and a compiled Numba engine proven to build the bit-identical graph. It is benchmarked against FAISS and hnswlib with one single-threaded method, and it matches FAISS's recall within 0.01 on three datasets. It also has filtered search, int8 quantization and memory-mapped index files. Benchmarks · Performance notes · Studies

two-stage-recsys. Two-tower retrieval served by hnsw-engine, then a LightGBM ranker. NDCG@10 of 0.0906 against 0.0445 for tuned ALS; 0.1047 at ten times the data; 0.56 ms per request. NDCG@10 on MovieLens-1M with full-catalog ranking: popularity 0.0193, item-item co-occurrence 0.0276, tuned implicit ALS 0.0445, two-tower with exact retrieval 0.0615, two-tower via hnsw-engine 0.0619, two-tower plus LightGBM ranker 0.0906.

two-stage-recsys is a movie recommender built the way large production systems are: a PyTorch two-tower model retrieves 200 candidates through hnsw-engine, a LightGBM LambdaRank model reorders them, and a FastAPI service returns the top ten. It is evaluated with full-catalog ranking, a leakage-tested temporal split and 95% bootstrap intervals. The first ranker scored below plain retrieval, and the write-up shows how that was diagnosed and fixed. Results · Analysis · Model card

options-pricing-engine. Black-Scholes, American trees, Monte Carlo and implied volatility, checked against QuantLib. Median relative gap to QuantLib prices 4e-16; all 18 Ju-Zhong American prices within 2e-3; 493 tests. Pricing error of an American put against runtime, log scales. CRR falls from 2.8e-2 to 1.3e-4 by 16 ms; Leisen-Reimer from 7.5e-3 to 1.5e-5 by 15 ms; BBSR, nearly tied with Leisen-Reimer up to about 10 ms, reaches 1.6e-6 at 206 ms.

options-pricing-engine prices European and American options with Black-Scholes-Merton, four binomial trees, Monte Carlo with variance reduction, and safeguarded implied-volatility solvers. Its results are checked against 50-digit arithmetic, against QuantLib on 1,800 cases, and against published American-option benchmarks. Its numerical studies had their hypotheses committed before they ran, and the write-ups report the ones that failed. One command, python -m verify, checks every number in its README against the file it came from. Results · Study A: implied volatility · Study B: trees

Research and experience

  • Pulse Energies, founder and CEO, 2025 to now. Physics-constrained AI for forecasting and orchestrating electricity grids.
  • SmartPyro, research. AI-Optimized Pyrolysis: A Machine Learning Framework for Predictive Waste-to-Energy Conversion, with Sanjith Balamurali, at the 3rd International Online Conference on Energies (IOCEN 2026). Also presented at ASME ES2026.
  • Lumiere Research Program, machine learning researcher, 2025. A PyTorch classifier for vector boson polarization in simulated Large Hadron Collider events, supervised by Dr. Gaetano Barone of Brown University.
  • Hubtel, technical product manager intern, 2026, on merchant payments. IPMC Ghana, software engineer intern, 2025, on an observability dashboard built with GraphQL and React.
  • Awards. IFoRE Medalist, first award in physics and engineering interdisciplinary research. SSRN Emerging Scholar Award.

Toolbox

Rust Python TypeScript NumPy SciPy Numba PyTorch LightGBM Tokio FastAPI Next.js PostgreSQL Docker pytest Java SQL

Pinned Loading

  1. hnsw-engine hnsw-engine Public

    HNSW vector search built in Python + Numba, benchmarked against FAISS and hnswlib up to SIFT-1M, with filtered search and int8 quantization.

    Python 1 1

  2. two-stage-recsys two-stage-recsys Public

    Two-stage recommender: two-tower retrieval served by hnsw-engine, then a LightGBM ranker. Doubles tuned ALS on MovieLens-1M and 10M (NDCG@10).

    Python 1