Projects

A working list of things I’ve built. Some are open source, some are client work, some are research code. They’re grouped by what they’re for rather than by what they’re written in.

Agent harnesses

The layer between a language model and software that already exists. My view: most teams give the model the wrong surface — raw tables, auto-generated specs — and then blame the model when it guesses.

Reins · Python

A lightweight harness you bolt onto an existing app so an LLM can operate it. Point it at functions you already have — or at your SQLAlchemy/Django models — and hand it a goal in plain language. It works out which operations to call and in what order, runs them through your own code, never raw SQL, and returns the result.

• ask() is read-only and cannot write; writes are opt-in and gated by policy, approval, and audit — enforced in code, not in a prompt the model can talk its way around.
• Agent.from_orm(...) introspects models into intent-named verbs (find_orders, get_user_by_email), preferring unique natural keys over internal IDs.
• Anthropic, OpenAI, Gemini, Groq, Z.ai, Ollama, and vLLM in the box; agent.explain() shows every call it made and why.

Veldra · Python · Vue 3

A self-hostable, local-first agent-harness platform. The load-bearing idea: an agent is data, not code — a versioned AgentSpec row in Postgres. “Build me an agent” compiles natural language into a validated spec; “change it” is a JSON-Patch you approve; the runtime is a pure interpreter of the current version.

• Natural-language agent and team construction (coordinator + specialists, depth-capped delegation).
• MCP connectors (Shopify, Alibaba, or any Streamable-HTTP / SSE / stdio server) alongside built-in tools; side-effecting connector tools require explicit approval.
• Editable knowledge bases with per-KB retrieval mode, embedding model, reranker, and vector store; citations carry page, section, and character span.
• A visual workflow builder, and agents that accumulate lessons from feedback and from their own failed tool calls.

Clawkido · Go

Actor-model multi-agent swarm engine: goroutine-per-agent routing, swarm handoff with depth-limited chains, an extensible skill system, provider fallback, and Telegram/Discord integration. Single binary, roughly 50 MB of RAM.

LangGraph Agent System · Python

Production-shaped multi-agent system on LangGraph Platform: specialized agents, multi-provider LLM support, LangGraph Studio compatibility.

Machine learning for biology

AIPpred-Stack — anti-inflammatory peptide prediction · IEEE CCECE 2026 with Abdullah Al Mamun and Francis M. Bui (University of Saskatchewan)

A two-stage stacking ensemble that combines ~2,282 hand-crafted sequence descriptors (AAC, DPC, DDE, CKSAAP, CTD, PAAC, QSO, autocorrelation, physicochemical) with 1,280-dimensional ESM-2 protein language model embeddings. The real contribution is not the architecture — it’s the dataset. Earlier methods drew negatives from random UniProt proteins, which quietly reduces the task to “short peptide vs. long protein” and inflates accuracy. We use experimentally validated negatives from the same IEDB T-cell assays as the positives and report the harder numbers that follow.

DeepNatProtease — natural-product inhibitors of viral proteases

An end-to-end virtual-screening pipeline: integrate per-virus bioactivity data (ChEMBL, BindingDB, PubChem BioAssay, ZINC, COVID Moonshot) with full provenance → standardize and harmonize units to nM → split by Bemis–Murcko scaffold clusters with zero overlap → train virus-specific ChemProp GNN ensembles and classical models → screen the COCONUT library of 400k+ natural products → filter PAINS/toxicophores and enforce scaffold diversity → dock survivors against ligand-informed structures → prioritize by consensus of ML probability, docking score, and drug-likeness. Strict per-virus isolation throughout: HIV-1 protease, HCV NS3/4A, SARS-CoV-2 Mpro, Dengue and Zika NS2B-NS3.

Sports science and biomechanics

Athlete Intelligence (AlgolyzerLab) · Go · Next.js 16 · Python

Multi-tenant athlete performance management for elite sport, built on an IOC-aligned clinical domain: athlete registry, injury and illness register, screening, fitness testing, workload, return-to-play, and dashboard analytics — with strict tenant isolation and a soft-delete-first data model so medical history is never destroyed. A separate GPU service runs RTMPose (Halpe-26) over ordinary video and turns it into clinical movement metrics using 2-D geometry and multi-view triangulation.

Cricket Bowling Analyzer (AlgolyzerLab) · Python · Go

Biomechanical analysis of a bowling action from a webcam, video, or a single image. MediaPipe gives 33 landmarks; twelve of them and some plain trigonometry give front-knee angle, elbow extension, trunk side-bend, and shoulder rotation — checked against injury-risk ranges and the ICC’s 15° elbow-extension rule. The Go version tracks 18 parameters at 30+ FPS, holds player profiles by bowler type, and generates PDF reports.

Agriculture AI

Agri Remedy · FastAPI · React 19 · TypeScript

Farmers upload a drone flyover; the app finds diseased patches and explains, in plain English or বাংলা, what the problem is and what to buy for it.

• DINOv2-Small as a frozen feature extractor with staff-trained linear-probe heads — the backbone supplies general visual understanding for any crop, and the crop-specific knowledge comes from labelled photos.
• Three-tier fallback so every crop works on day one: trained linear probe → k-NN centroid gallery → zero-shot colour/texture heuristic with confidence capped well below a trained model.
• Canopy and lesion segmentation drives severity (affected canopy area), the snapshot crop, and the anomaly mask; ISO-6709 GPS tags read from video metadata via ffprobe pin findings on a representative aerial frame.
• A staff console for the disease library, the product catalog, and the per-crop models — label boxes on real photos, train, and read a genuine cross-validated accuracy.

Healthcare software

ChemberIQ (AlgolyzerLab) · Go · Vue 3 · Flutter

A medical case-management platform for doctors and clinics: public doctor portfolio, secure clinical CRM, AI chatbot, real-time doctor↔patient WebSocket chat, and an offline-first Flutter doctor app. The doctor types their clinic website, the app discovers the backend, downloads the dataset, and works fully offline — with a big-text, big-button UI designed for older doctors. Search is a purpose-built in-memory engine (prefix → word → contains → token → fuzzy) that replaced a Meilisearch dependency entirely.

Systems and developer tools

Fennec · Go

Zero-dependency library for SSIM/MS-SSIM-guided perceptual image compression: quality targeting, batch worker pools, auto format selection, Lanczos-3 resize, EXIF auto-orientation, a target-file-size engine, and context-aware cancellation. 60–90% size reduction at SSIM ≥ 0.94.

VoidMon · Go

Terminal system monitor with per-core CPU, cross-platform GPU support (NVIDIA via nvidia-smi, AMD via ROCm, Intel via sysfs, Apple Silicon via powermetrics), disk I/O, network throughput, battery, and top-process tracking.

RAG-Scout · Python

Benchmarks sparse, dense, hybrid, late-interaction, and reranked retrieval stacks on any Q/A dataset, then tells you which one to actually build on.

LexSubLM-Lite · Python

Laptop-friendly lexical substitution: prompted generation via 4-bit causal LLMs, POS and morphological filtering (spaCy + pymorphy3), log-prob and cosine ranking, research-grade metrics (P@1, Recall@k, GAP, ProF1), and a YAML registry for swapping models without code.

VibeVoice Studio · Python

Web app over Microsoft’s VibeVoice TTS: train a voice from uploaded or recorded audio, multi-speaker synthesis up to four speakers, cloning, real-time audio visualization, and a voice library.

PlugBot · Next.js · FastAPI

Dashboard for running many Dify apps and Telegram bots at once: streaming replies, health checks, encrypted secrets, one-command Docker deploy.

Dify plugins

Published for the open-source Dify community:

• GigaChat model provider — Sber GigaChat chat, vision, function calling, and embeddings with configurable scopes and SSL handling.
• PDF → image converter — page rasterization with DPI and quality options, file URLs, batch processing.
• Nmap network scanner — port, OS, service, and vulnerability scans with safe-mode defaults and JSON/XML output.
• Together AI image and FLUX Fill Pro — text-to-image and professional inpainting/outpainting.
• Telegram and Slack integrations.

The full list, including the experiments that didn’t work out, lives on GitHub.