LALITH ALPURI · AI SYSTEMS ENGINEER
I design and ship local-first AI infrastructure—from agent memory and model tooling to multi-agent simulation—and make each layer inspectable.
Local models. Persistent memory. Agent worlds.
Designed to run close, stay clear, and earn trust.
Good agent systems carry decisions, evidence and working context beyond a single conversation.
Models and data stay close to the machine. The system boundaries stay visible.
Memory, tools and simulation earn their place when they support real work and decisions.
Each project explores a different layer of autonomous software. Select a system to see what it actually does.
MODEL ROUTING / MLX / LOCAL TOOLING
A Python runner routes prompts between a 7B verifier and 1.5B draft path, a 135M reflex model, targeted web lookup and an optional code route.
The current router is keyword-based. The source and local checkpoints are real; each execution path still needs its own runtime validation.
A 4-bit Qwen2-family UWOF 7B v3 checkpoint is present on disk. MLX LoRA scripts prepare JSONL data for local fine-tuning, and a separate 7B adapter checkpoint exists.
The saved adapter was configured for 28 adapted layers, rank 8 and 100 iterations on a Qwen2.5 7B 4-bit base. A fused 8.5B release is not established by the inspected artifacts.
The organism includes a web-search tool for targeted research. toast1_fetch() searches Hacker News through Algolia and retrieves company information for research prompts, then passes useful source text into the local answer workflow.
run_autonomous_agent() routes coding tasks to OpenInterpreter with an Ollama-hosted UWOF model. The integrated execution path lets the agent write and run code as part of its workflow.
The active MLX inference stack routes work across the UWOF 7B model, Toast 1.5B draft model and a 135M reflex model. Keyword-based routing selects the tier suited to each request, keeping generation on the local Apple Silicon setup.
Technical work spanning AI-assisted simulation, market analysis, predictive modeling and competitor intelligence.
Worked with a million-row Monte Carlo dataset in SQLite. Evaluated pricing, commissions, landed costs, net selling price and margins through feasibility and profitability filters; produced ranked exports and an HTML explorer.
Modeled acquisition cost, conversion, ROAS and channel assumptions. Used synthetic consumer scenarios to identify validation priorities rather than claim unmeasured outcomes.
Structured comparisons and AI-assisted evidence briefs to connect source material, assumptions and limitations to product-planning decisions.
MLX · LoRA · Ollama · MCP · RAG · agent workflows
Python · Java · SQL · SQLite · REST APIs · Spring Boot
JavaScript · React · WebGL · Docker · GitHub Actions
Monte Carlo simulation · predictive analysis · market intelligence
Open to serious engineering conversations.