Selected work.

A few examples across civic data, open-data agents, sports vision, education, creative tools, and local infrastructure.

How we work with teams.

network_node

Agentic engineering

Agents with review, evals, and handoff

We use agents as part of an engineering workflow: clear briefs, scoped tasks, evals, tool permissions, review gates, and handoff notes. The goal is not more automation for its own sake. It is better work the team can understand and keep.

Claude CodeCodex CLIGemini CLISnowflake CocoParallel agentsEvals-firstGovernance

hub

AI architecture

Patterns for regulated teams

We help teams design tenant isolation, model routing, retrieval, audit trails, prompt governance, and hybrid deployment paths without turning every use case into a new platform.

Tenant isolationModel routingPer-tenant RAGAudit trailsHybrid cloudOn-prem

account_tree

Shared operating model

Connect the tools to the workflow

Teams usually have scattered AI tools before they have a shared operating model. We help connect data, workflows, policies, and decisions so AI work becomes easier to review, govern, and reuse.

AI strategyOrg-level RAGKnowledge graphsDecision systemsAI roadmap

groups

Team adoption

Team habits for agent-assisted work

Adoption works when people know what changes in the day-to-day. We help teams define review habits, eval ownership, prompt and model governance, and the standards for work produced with agents.

Org designMLOpsLLMOpsEvalsCareer laddersGovernance

memory

Small + Local LLMs

Local models where they make sense

For private workflows, small open models can reduce cost, keep data close, and give teams more control. We help decide where local models fit and how to serve them cleanly.

OllamaLM StudioUnslothLlama 3Kimi K2DeepSeek V3Nemotron OmniFine-tune on private data

See the full stack arrow_forward

Projects.

Built with a mix of engineering, local models, cloud models, and agent-assisted workflows.

Civic Intelligence

1st place · NVIDIA DGX AITX

Undervolt

Urban growth intelligence that reads 2.2M Austin building permits and surfaces where the grid is about to break — running on a 60W Jetson at the edge.

Nemotron Nano 8BJetson AGX OrinRAPIDSDeepStreamNext.jsSupabase

2.2M

Permits

Params on edge

22:1

Solar-battery gap

60W

Inference power

Problem

Austin permits are public but useless — 2.2M rows of messy fields no one reads. Energy planners, equity advocates, and city staff all miss signals that should drive policy.

Approach

LLM + RAPIDS pipeline that structures every permit, links it to weather events, and runs Nemotron on a Jetson so the whole thing runs at 60W — no cloud bill, no data egress.

Outcome

Surfaced a 22:1 solar-to-battery install gap and a 12× equity gap in backup power. Took 1st at NVIDIA DGX AITX. Presented to City of Austin planning staff.

Read the case study arrow_forward

Civic Data

Open source

TXLookup

A Texas open-data agent that turns plain-English civic questions into sourced answers across Austin, Dallas, San Antonio, Houston, and state portals.

6,061 datasets7 agentsMCP toolsSocrataCKAN

Problem

Texas public data is spread across thousands of datasets and portals. Useful answers require schema reading, query writing, and source checking.

Approach

A multi-agent workflow picks datasets, writes bounded SoQL queries, runs them against source portals, and attaches citations to every answer.

Outcome

6,061 datasets indexed, 8 MCP tools exposed, and replayable answer traces for civic questions across major Texas open-data portals.

Launch app open_in_new View on GitHub open_in_new

Sports AI

Live product

RefereAI

A personal AI coach, referee, and guide for 240M tennis matches a year that never get a line-call review. VLMs reason about trajectory, spin, and bounce — then explain the call.

Cosmos Reason 2DeepStream 7.1Jetson AGX OrinFastAPINext.js

Problem

Rec and youth tennis has no reliable officiating. Disputed calls end games. Coaches can't be in every match. Video review is too expensive for non-pros.

Approach

Cosmos Reason 2 VLM running on Jetson at the edge. The model watches the ball, reasons about physics with chain-of-thought, and narrates each call in plain language.

Outcome

Live on-device inference at 60W. Chain-of-thought explanations parents and players actually understand. Zero-cloud privacy — video never leaves the court.

Launch app open_in_new Read the case study arrow_forward

Physical AI

Live product

Sideline

Agentic AI sports referee that doesn't just watch the game — it reasons about the play and moves a robot arm to signal the call. Open source.

Qwen VLMJetson AGX OrinSO-101 ArmUnitree G1Three.jsWebSocket

Problem

Youth and rec sports run on volunteers with no officiating budget. Existing AI refs stop at a verdict on a screen — no way to communicate the call to the field.

Approach

Full agentic pipeline: VLM watches video → chain-of-thought reasons about the play → agent issues commands → physical robot embodies the call. Multi-provider VLM switching.

Outcome

Working prototype at Nebius Build SF. Three embodiments wired in (arm, humanoid, rover). Open-source codebase so leagues can run it themselves.

Launch app open_in_new Read the case study arrow_forward

Education

Live product

StudyPal

The homework buddy that asks better questions. Voice tutoring, Schoology sync, and adaptive practice — built to teach, not hand over the answer.

Gemini LiveGemini VisionVoice WebSocketSchoology APISupabaseCloud Run

Problem

Homework AI tools either refuse to help or give the answer. Neither teaches. Parents want a tutor that guides their kid through the work — and proves it happened.

Approach

Voice-first tutor on Gemini Live. Vision grades journal photos. Auto-syncs with Schoology so assignments are already loaded. Adaptive practice fills the gaps.

Outcome

In daily use with real students. Parent Telegram digests summarize what the kid worked on. PIN-protected family accounts — privacy-first by default.

Launch app open_in_new Talk to us about StudyPal arrow_forward

Biotech · Drug Discovery

Live product

HD Research Hub

Open-source AI research platform for Huntington's Disease drug discovery. LLM agents triage biomedical literature, surface repurposing candidates, and update themselves daily — so families and researchers aren't reading 40-year-old summaries.

OllamaGeminiPubMedClinicalTrials.govHDBuzzOpen TargetsAlphaFoldVercel

Problem

HD families and clinicians navigate a research landscape that moves faster than any one specialist can read. Most decision-support tools are paywalled, static, or pharma-owned.

Approach

Four autonomous agents pull from PubMed, ClinicalTrials.gov, HDBuzz, and Open Targets. Ollama-backed LLM triage ranks relevance, generates drug-repurposing hypotheses, and cites every claim. Refreshes daily via GitHub Actions.

Outcome

Live at hd-research-agent.vercel.app. Public dashboard, experiment write-ups, and target landscape — all open source. Runs on Jetson or Mac, no cloud API required.

Launch app open_in_new View on GitHub open_in_new

Other builds worth knowing about.

Additional builds across creative tools, youth sports, job search, and infrastructure.

brush

Creative AI

Studio Copilot

Local AI workspace for photographers and designers. Full workflow — curation, client review, contracts, invoices, feedback — all on-device. Runner-up at GTC Hack to Create. Featured on NVIDIA Developer Live.

OpenCLIPNVIDIA GB10FastAPINext.js

View project arrow_forward

sports

Youth Sports

CoachClaw

AI sideline assistant for volunteer parent-coaches. Rules lookups, concussion protocol, practice plans, park safety, restaurant health scores — all from a Telegram message. Zero cloud dependency.

NemotronNVIDIA GB10Telegram Botllama.cpp

View project arrow_forward

work

Job Search

Applly

AI-powered job search copilot. Scores resume-to-job fit, generates tailored cover letters, and prepares interview talking points. Chrome extension imports listings directly from LinkedIn.

GeminiNext.jsChrome ExtensionSupabase

hub

Infrastructure

Homenest

Multi-device observability dashboard. Monitors services, GPU utilization, network topology, and AI model health across Mac and Jetson nodes — with Google Analytics integration and Telegram alerting.

FastAPIWebSocketsGoogle AnalyticsTelegram

Advisory work.

AI Retail Platform

AI-powered style discovery for home furnishing retailers. Increased average order value by 66% through automated style reasoning.

+66% AOV

Water AI Intelligence

Real-time digital AI infrastructure for 120+ water utilities. ATLAS, AURA, APOLLO platforms for predictive maintenance, leak detection, and treatment optimization.

120+ utilities

Technical Advisor

Advised a global personal data exchange on platform architecture, API design, and engineering strategy. Privacy-first data monetization.

Privacy-first

Advisory details arrow_forward

Have an AI project in mind?

Send the context. We will help shape the next step.

Contact AISOFT arrow_forward