I build AI that finds the right source, acts on it and gets checked.
Hybrid RAG & reranking · approval-gated MCP agents · LLM-as-judge evals · realtime voice agents
- RAG and retrieval. Hybrid dense + sparse vector search across several indexes, multi-query expansion with reciprocal rank fusion, one global Cohere rerank, and answers grounded only in retrieved passages. Scored with a RAG evaluator, not eyeballed.
- Agents with a human in the loop. MCP and native tool calling, default-deny tool policies, hard step and dollar caps. Reads run, writes wait for a click. Plus a terminal coding agent.
- Evals, not vibes. LLM-as-judge scoring, generated test suites for voice agents, human expert scoring side by side.
- Voice, in real time. Speech-to-speech agents on LiveKit and OpenAI Realtime that call retrieval tools mid-conversation, plus a phone receptionist on a real Asterisk line.
- The model drafts, code does the math. Structured outputs, deterministic parsers, every value traced back to its source. Multilingual, including Persian (RTL) support.
Capabilities, and the repos that prove them
| Capability | Proof |
|---|---|
RAG, retrieval & vector search (hybrid dense + sparse over LlamaCloud vector indexes, text-embedding-3-large, multi-query fusion, Cohere rerank, grounded answers, RAG evaluation) |
rag-service · rag-evaluator · livekit-fa-agent · product-assistant |
| Realtime voice agents (LiveKit, speech-to-speech) | livekit-fa-agent · voice-chess-coach · product-assistant |
| Telephony voice AI (Asterisk, Android dialer) | phone-agent |
| Tool-calling / MCP agents (approval gates in pm-assistant, pr-agent, dongeto) | pm-assistant · agent-studio · pr-agent · dongeto |
| MCP servers (14 on PyPI, Iranian services) | snapp-mcp · digikala-mcp · mrbilit-mcp · bale-mcp · payping-mcp · and 9 more |
| Coding agents | sepicode · pr-agent |
| LLM evals (LLM-as-judge, test generation, human scoring) | voice-agent-testgen · rag-evaluator · polymind |
| Structured output and document AI (OCR, diarization, vision) | invoice-extractor · meeting-assistant · feedo · ai-explain · steel-market-analyst |
| Text-to-SQL with guardrails | erp-sql-agent |
| AI safety (DLP, sandboxing, prompt-injection guards, step and spend caps) | polymind · invoice-extractor · ai-explain · sepicode |
A hybrid retrieval API: one HTTP call turns a question into a short, reranked list of source passages that any LLM agent can use as a retrieval tool.
Hybrid dense + sparse search over several LlamaCloud vector indexes ( text-embedding-3-large) · 3 LLM paraphrases + reciprocal rank fusion · one global Cohere rerank, falls back to fusion order on failurePython · FastAPI · LlamaIndex · LlamaCloud · Cohere Rerank |
A local-first AI copilot for project managers: reads and acts across Jira, GitHub, Slack and more, and every write waits for your click.
Default-deny tool policy · approve/reject mid-stream (browser or Telegram) · plain-English rules compiled once, polled with zero LLM calls Python · FastAPI · React 19 · SQLite · MCP |
A self-hosted, multi-model AI workspace with a DLP gate, sandboxed code execution and first-class Persian (RTL) support.
Side-by-side arena + judged debates between 2-5 models · DLP scan before any provider · LLM-written pandas runs under bubblewrap, no network Python · FastAPI · PostgreSQL · React · OpenRouter |
A Persian voice agent that talks back in real time and answers only from retrieved passages, built to work on filtered networks.
RAG via a search_knowledge tool call · turn-taking tuned for Farsi pauses · UDP → ICE/TCP fallbackPython · LiveKit Agents · OpenAI Realtime · nginx |
A full-screen terminal coding agent that reads, edits, searches and runs your code, with hard step and dollar caps on every turn.
Streaming Markdown, live tool-call timeline, inline diffs · nine zod-typed tools · read-only sub-agent capped at 8 steps / $0.25 Bun · TypeScript · React · OpenTUI · OpenRouter |
Drop in a Persian meeting recording; get back who said what, the decisions, the action items and a follow-up email.
Parallel chunked transcription with speaker stitching across seams · minutes from diarized words, summaries via strict JSON schema Python · FastAPI · Next.js · ElevenLabs Scribe v2 |
Plug Iran's everyday apps into Claude, Cursor, VS Code or any MCP client. Each one is a small Python stdio server on PyPI: uvx <name> and go.
Search Snappfood and SnappMarket food and groceries, compare full prices with packaging and delivery, find deals.
|
Search Digikala with feature filters, best picks for a budget, every seller's price, price history and reviews.
|
| mrbilit-mcp Travel · cheapest flights, trains, buses and hotels, with seats and refund rules |
jabama-mcp Travel · villas and stays at exact prices, tours, events and theater tickets |
otaghak-mcp Travel · villas and cottages, availability, exact prices, reviews |
| technolife-mcp Shopping · electronics, sellers and installment prices |
masterkala-mcp Shopping · gadgets and accessories, prices and specs |
shopino-mcp Fashion · thousands of shops, sizes, stock, shop ratings |
| khanoumi-mcp Beauty · cosmetics and perfume, shades, reviews |
asalbanoo-mcp Beauty · skin and hair care, variants, deals |
doctoreto-mcp Health · doctors, visit fees, free slots, clinics, labs |
| paziresh24-mcp Health · doctor profiles, prices, reviews, free slots |
bale-mcp Messaging · your Bale bot reads, answers and sends files |
payping-mcp Payments · your PayPing sales, customers, payment links, invoices |
Store, travel and health servers are unofficial and read-only. Bale and PayPing run locally with your own token.
 product-assistant Scan a product, ask it out loud about sugar or allergens. Realtime voice grounded in pack facts. |
 phone-agent Persian AI phone receptionist on a real line: Asterisk AudioSocket, 8 kHz PCM, plus an Android dialer app. |
 voice-agent-testgen Paste a voice agent's prompt, get a judged test suite back. Runs the real agent via LiveKit's AgentSession.
|
 agent-studio Persian-first AI workbench: native tools and MCP servers (stdio, HTTP, SSE), every step visible in the chat. |
 pr-agent Seven AI reviewers on every GitLab merge request, one human who decides what gets posted. |
 erp-sql-agent Ask your ledger in Persian, get audited numbers. SELECT-only SQL tool with a forced LIMIT.
|
 rag-evaluator LLM-as-judge on four criteria, plus human experts scoring the same answers with AI scores hidden. |
 steel-market-analyst Price PDFs in, grounded market read out. 13 named hallucination failure modes in the prompt pack. |
 invoice-extractor OCR + LLM extraction, every value linked to its box on the page. Injection-guarded, runs fully local. |
 dongeto Type the expense in Farsi. The LLM drafts it, unit-tested code does the money math. |
 feedo Plain-Persian bug reports plus screenshots in, structured drafts out. The model never picks IDs. |
 ai-explain Ask anything, get an interactive HTML/SVG explainer, planned and streamed into a sandboxed iframe. |
 voice-chess-coach Play White by voice against a coach on a live board. tool_choice="required" + python-chess own the rules.
|
Off the clock: True Anomaly, a WebGPU space sim
Fly a 6DOF probe through a real-scale solar system in the browser, every planet placed from NASA/JPL Horizons data. Physics in true kilometres with a Newton-iteration Kepler solver, rendered on Three.js
WebGPURenderer with a WebGL fallback.
All screenshots use demo or synthetic data. Open any repo for architecture notes, setup and more screenshots.
My last year of commits as Breakout bricks, regenerated daily. Play it yourself · made with pacman-contribution-graph