NumunInformation
Technology
Selected work

Four engagements.
Four different problems.

Research infrastructure, cryptographic trust, classrooms and commerce. Client names are withheld where agreements require it; the work, the stack and the outcomes are as delivered.

IT & software · Applied AI

An expert-reasoning data platform for a frontier AI research lab

Client: an applied AI research lab, United States. Name withheld under NDA.

  • 4
    data product lines launched
  • 12 wk
    from brief to first paid delivery
  • 3×
    contributor throughput after workflow redesign

Situation

The lab had a strong thesis: models trained on expert reasoning improve where models trained on outputs plateau. What it lacked was the machinery to capture that reasoning from professionals at scale, grade it consistently and package it for customers building foundation models.

What we built

A contributor platform with domain-specific task templates for supervised fine-tuning pairs, rubric-graded reinforcement learning prompts, tool-based agent environments and recorded computer-use trajectories. A two-tier review pipeline with automated consistency checks, and a delivery layer that exported datasets in each customer's schema.

Outcome

Four product lines went live in twelve weeks. Reviewer agreement rose above the customer's acceptance threshold in the second sprint, and contributor throughput roughly tripled once the task UI stopped fighting the experts. The platform now carries the lab's enterprise deliveries.

PythonTypeScriptPostgresLLM evaluationRubric designData pipelines
IT & software · Trust infrastructure

A post-quantum proof layer for credentials and documents

Client: a European trust-infrastructure startup. Name withheld.

  • <1 s
    browser-side verification
  • 2,000
    records anchored per batch
  • 0
    source documents stored on-chain

Situation

Regulatory deadlines on AI-content disclosure and post-quantum cryptography were arriving faster than the client's customers could re-platform. The founders wanted proofs that anyone could verify in their own browser without trusting the vendor, at a cost per record that made bulk credentialing viable.

What we built

A four-step pipeline: SHA-256 fingerprinting, Merkle batching anchored to a public chain, hybrid signatures combining Ed25519 with a NIST post-quantum scheme, and a stand-alone browser verifier that recomputes everything client-side. Around it, a REST API for hashing files or JSON records, branded verification pages, and an issuer console for credentialing teams.

Outcome

Verification completes in under a second with nothing sensitive leaving the customer's environment, which kept the design GDPR-compatible by construction. Batched anchoring brought the marginal cost per record to a fraction of a cent, and the client launched with two products and a roadmap into content provenance and product passports.

RustNode.jsPolygonML-DSAEd25519REST APIWebCrypto
Education · AI in schools

A district-ready AI assistant platform for K-12 teachers

Client: an education technology company serving school districts. Name withheld.

  • 80+
    teacher tools at launch
  • 7–10 h
    reported weekly time saved per teacher
  • SOC 2
    controls designed in from day one

Situation

Teachers were already using consumer AI tools, without guardrails, without district oversight and without any link to the curriculum. The client wanted a platform districts could adopt with confidence: safe for students, aligned to standards, and integrated with the systems schools already run on.

What we built

A tool library for lesson planning, rubrics, differentiated materials, feedback and assessment, alongside a student-facing layer that teachers launch and monitor. Integrations with the major classroom and LMS platforms, district dashboards for usage and governance, privacy architecture that never trains on student data, and a professional-development track with certification.

Outcome

The platform launched with more than eighty teacher tools and a district administration console. Pilot teachers reported seven to ten hours saved each week, and the governance layer was what moved conversations from individual teachers to district-level agreements.

Next.jsLLM orchestrationLTI / LMS integrationFERPA-aligned designLearning analytics
Marketing technology · Social commerce

Comment-to-conversation automation that proves which posts pay

Client: a social commerce SaaS for creators and online shops. Name withheld.

  • ~1 s
    from comment to tracked DM
  • 100%
    of commenters captured, 24/7
  • CTR
    reported per post, per campaign

Situation

Creators and shops were answering hundreds of comments by hand, losing most of the intent in the process and never learning which post actually produced sales. Existing tools stopped at auto-replies; nobody closed the loop to revenue.

What we built

Real-time comment monitoring via the official platform API with polling as a fallback, keyword and story-reply triggers, an instant direct-message flow with optional follow gates, a lightweight CRM logging every commenter's history, and tracked links reporting sends, clicks and click-through rate per campaign. Public replies post separately so the customer is reached even when a DM fails.

Outcome

Every commenter is now captured and answered within about a second, and for the first time the client's users could see which reel paid for itself. A free tier drives adoption; the paid tier unlocks unlimited campaigns and AI replies for unmatched questions.

Meta Graph APIWebhooksNode.jsPostgresAttributionGrowth analytics
Your project

What should the fifth one be?

If a problem above looks like yours, say so. If none of them do, that is usually the more interesting brief.

Start a projectEmail us