Experimental R&D

The Lab: Not everything starts as a product.

The Lab is where I test hypotheses, benchmark new AI models, experiment with on-device inference, and document raw discoveries before packaging them into software.

01
Hypothesis FirstClear testable thesis before any code.
02
Timeboxed Sprints7-day prototypes to prove or kill ideas.
03
Document FindingsShare what worked and what broke.
#014Autonomous Agent
In Progress

Autonomous Niche Opportunity Scraper & Viability Classifier

1. Hypothesis

Can an autonomous AI agent scrape unranked marketplace niches, analyze 1-star review clusters, and validate search velocity to propose viable digital micro-products in under 5 minutes without human intervention?

2. Method / Process

Constructed a 4-node LangGraph agent with Playwright headless scraper, Cohere embedding clustering, and Gemini 2.0 Flash unit economics evaluator. Ran across 40 sub-categories in utility software and mobile productivity.

3. Result

Agent successfully isolated 12 high-intent unmet software features with under 3 competitive solutions. Identified recurring complaints around pet health coordination and niche technical exam preparation.

4. What I Learned

Pure keyword volume is a lagging indicator. The highest signal lies in negative review sentiment clusters with phrases like 'I wish this app also had...' or 'Impossible to do X with my spouse'.

Run Time:3m 42s
Accuracy of Niche Filter:89%
Review Density Scanned:24,000+
#Agent Loops#E-Commerce#Market Discovery#LangGraph
#011Latency & Voice
Completed

Sub-200ms Latency Audio Pipeline for Real-Time Language Coaching

1. Hypothesis

Can we achieve natural, conversational turn-taking (< 200ms latency) in Italian language voice coaching using WebSockets, streaming Whisper STT, and small quantized local TTS models?

2. Method / Process

Benchmarked direct client WebSocket streaming vs Server-Sent Events with edge proxying. Tested OpenAI Realtime API vs modular pipeline (Whisper Edge -> Groq LLaMA-3-8B -> Kokoro TTS).

3. Result

The modular edge pipeline achieved 230ms average glass-to-glass latency, while OpenAI Realtime API hit 180ms. Kokoro TTS provided natural Italian cadence with minimal server footprint.

4. What I Learned

Voice coaching doesn't just need fast latency; it requires conversational backchanneling ('um', 'certo', 'esatto') to prevent students from feeling awkward while thinking in a foreign language.

Avg Latency:184ms
Natural Cadence Score:9.1/10
#Web Audio#Voice AI#Low Latency#Whisper#Kokoro TTS
#008Prompt Optimization
Completed

Dynamic Prompt Reranking vs. Fine-Tuning for Bureaucratic Legal Text

1. Hypothesis

Does dynamic few-shot retrieval (RAG with semantic similarity) outperform an instruction fine-tuned 8B model when explaining tricky Italian highway legal rules?

2. Method / Process

Created a benchmark test suite of 400 'trucchi e tranelli' (trick questions) from the Italian Ministry of Transport. Evaluated a fine-tuned Mistral-7B against GPT-4o with dynamic vector retrieval of Codice della Strada articles.

3. Result

Dynamic few-shot RAG achieved 96.2% factual legal accuracy versus 84.1% for the fine-tuned model, while eliminating fine-tuning maintenance when legal amendments are passed.

4. What I Learned

For regulatory domains where laws change and exact wording matters, RAG + strong reasoning models consistently defeats fine-tuning, which suffers from subtle parametric hallucinations.

RAG Accuracy:96.2%
Fine-Tuning Accuracy:84.1%
#RAG#Legal AI#Evaluation#Fine-Tuning
#005Local SLM
Concept

On-Device SLM (Small Language Model) for Offline Health Symptom Logging

1. Hypothesis

Can a 1.5B parameter quantized model (like Qwen2.5-1.5B or Gemma 2 2B) run on mid-range Android hardware to parse voice symptoms into structured health JSON offline?

2. Method / Process

Tested ONNX Runtime and MediaPipe GenAI on Snapdragon 7-series and Tensor G2 chips with 4-bit INT4 quantization. Measured battery drain and memory footprint.

3. Result

Gemma 2 2B INT4 ran at 14 tokens/sec with ~650MB RAM footprint. Capable of parsing free-form symptom descriptions into structured JSON tags without internet connectivity.

4. What I Learned

Mobile hardware is now fast enough for zero-cloud, privacy-first AI apps. Pet health and personal notes can stay 100% on the device.

Tokens / Sec:14.2
Memory Footprint:640 MB
#Edge AI#On-Device#Privacy#Quantization#Mobile
#002Architecture
Completed

Adversarial Multi-Agent Debate for Hallucination Suppression

1. Hypothesis

Does a two-agent adversarial debate (Generator vs Skeptic) significantly reduce hallucination rates in technical code generation compared to single-pass CoT?

2. Method / Process

Set up 2 agents: Agent A produces code and explanations; Agent B acts as a senior reviewer with strict linter and test case execution tools. Evaluated on 100 tricky edge-case programming briefs.

3. Result

Syntax and logic regressions dropped by 73%. Agent B caught 91% of edge cases (null checks, race conditions) before output delivery.

4. What I Learned

Agents need separated roles with asymmetric incentives. Giving the reviewer agent a 'bounty' for finding bugs produces far better results than asking a single agent to 'review its own work'.

Regression Drop:-73%
Edge Case Catch Rate:91%
#Multi-Agent#Code AI#Adversarial#Reliability