Projects

AI Technical Due Diligence Assistant

Problem

Before capital is committed, a deal team has to judge a target’s technology from whatever exists (technical and non-technical, often incomplete), and the target frequently doesn’t know what it doesn’t know about AI. Reading it all and turning it into a defensible view is slow, and generic AI tools produce vague memos that don’t hold up in the investment committee.

Use case

PE and family-office deal teams and technical due-diligence leads running an AI roll-up / buy-and-build strategy, plus the operating partners who will own value creation post-close.

Demo

1. Ingest the target’s documents (architecture notes, GitHub summary, product docs, website text, commercials), labeled technical or non-technical.
2. Each document is chunked, embedded, and stored in a per-company knowledge base (RAG).
3. For each diligence dimension, the most relevant passages are retrieved and passed to Claude as grounded context.
4. Claude returns a structured memo whose findings cite the evidence, with the gaps turned into follow-up questions.

Output

Illustrative output. The memo also covers strengths and weaknesses, the best AI path forward, and roll-up fit. An early-stage screen, not a substitute for confirmatory diligence.

Tech stack

• Next.js
• TypeScript
• Claude (Opus 4.8)
• RAG
• Postgres
• pgvector
• Prisma
• Voyage embeddings

Business value

• Turns a pile of mixed target information into a defensible, IC-ready diligence view in minutes.
• Grounds every finding in the target’s own documents, so the memo is specific rather than generic.
• Surfaces the best AI path forward (value creation) alongside the risks, in one pass.
• Assesses roll-up fit: platform vs bolt-on, integration risk, and how cleanly the stack consolidates across a portfolio.

View on GitHub

AI Value Creation Roadmap Generator

Problem

After acquiring a business, the operating team has to answer one question fast: where does AI actually create value here? Generic AI-strategy decks don’t map to the specific company, and manually auditing every function for high-ROI opportunities is slow, subjective, and hard to compare across a portfolio.

Use case

PE operating partners and deal teams, portfolio-company CEOs and CTOs, and the value-creation team building the post-close AI roadmap.

Demo

1. Enter the portfolio company’s profile: industry, team size, pain points, current tools, data, budget, and risk tolerance.
2. An LLM agent extracts signals and screens candidate AI workflows across each function.
3. It ranks the top opportunities by impact, difficulty, and risk, keeping a human-in-the-loop check on sensitive steps.
4. It outputs a prioritized roadmap, an ROI score per initiative, success metrics, and a “do not automate yet” list.

Output

Illustrative output. Scores are a prioritization signal for the operating team, ranked by an ROI model.

Tech stack

• Next.js
• TypeScript
• React
• OpenAI (GPT-4o)
• Agentic workflow
• Structured JSON schema
• ROI scoring engine
• Tailwind CSS

Business value

• Turns a multi-week AI discovery exercise into a same-day first-pass roadmap.
• Standardizes how AI opportunities are found and prioritized across the portfolio.
• Gives the investment committee and portfolio leadership a defensible, ROI-ranked plan tied to value levers.
• Builds the first-90-days PMI AI roadmap with human-in-the-loop guardrails on sensitive workflows.

SafeTracks: AI Platform-Edge Safety Announcements

Problem

On crowded train platforms, passengers drift over the edge, rush the doors, or stumble near the tracks, and a moment’s inattention can be fatal. Generic alarms and static signage get tuned out, while the operator carries the safety liability and the service disruption of every incident.

Use case

Rail and transit operators (safety and operations teams), station managers, and the infrastructure operators and investors who own transport assets.

Demo

1. A platform camera frame is captured from existing station CCTV.
2. A compact on-device vision model flags unsafe behavior (leaning over the edge, crossing the yellow line, entering the track area) with a severity level.
3. The system writes a short announcement addressed to the specific person and location.
4. It speaks it in a familiar, attention-catching station voice (Yamanote-line style) in real time, so the warning actually lands.

Output

Illustrative output. Announcements are targeted to the person and location, and spoken in a familiar station voice to catch attention.

Tech stack

• Python
• Liquid AI LFM2.5-VL (vision)
• Fine-tuned Liquid audio TTS
• PyTorch
• Transformers
• On-device / edge inference

Business value

• Turns existing platform cameras into a real-time fall-prevention system.
• Person-specific announcements in a familiar voice cut through alarm fatigue, so warnings change behavior in the seconds that matter.
• Runs on small, efficient on-device models, keeping cost and latency low and keeping video private (no cloud).
• Reduces accident liability and service disruption for transit operators.

IP Risk Screening Tool for Technical Due Diligence

Problem

Undisclosed IP exposure can surface after close and derail the value-creation plan. Running formal legal review across every product feature and patent claim is slow and expensive, so teams need a way to triage where the real exposure is.

Use case

PE deal teams and technical due-diligence leads, with IP counsel as the downstream reviewer, and portfolio-company CTOs.

Demo

1. Load the target’s product features and a body of relevant patents.
2. Match each product feature to the most relevant patent claims.
3. Score similarity and assign a triage risk level.
4. Output a prioritized review list and save the report for the diligence trail.

Output

Illustrative output. Similarity scores are a triage signal, not a measure of legal infringement.

Tech stack

• Python
• Flask
• Fuzzy claim matching
• LLM scoring path
• Docker

Business value

• Surfaces IP red flags early, before expensive legal review begins.
• Directs counsel to the highest-risk items first.
• Feeds deal pricing and risk assessment during diligence.
• Leaves an auditable trail of what was screened and flagged.