Engineering Signals – Akshit

I do not trust AI output until it is evaluated.

I am building toward systems where generated code, extracted data, and model responses are tested, scored, and validated before humans rely on them.

I build AI that survives messy data.

At BlackRock, I worked on validation-first pipelines for reports, holdings, trades, and extracted portfolio signals instead of assuming clean inputs.

Code generation is only step one.

TrustLoop turns prompt, code, and GitHub changes into generate, red-team, execute, evaluate, and repair loops with evidence attached to every run.

I think in failure cases first.

I like systems where an agent writes the happy path and another one attacks edge cases, security assumptions, and performance cliffs.

Finance-grade AI needs boring excellence.

Latency, auditability, deterministic outputs, data contracts, and reproducible validation matter more than flashy demos in enterprise finance workflows.

Dashboards should shorten decisions.

The Fideuram dashboard consolidated positions, trades, exposure deltas, and allocated capital so analysts could reconcile faster with less manual stitching.

Documents are data pipelines in disguise.

I used OCR, PDF parsing, NLP, and deterministic extraction to turn Quarterly Portfolio Reviews into structured signals for downstream agent workflows.

Retrieval quality has a speed limit too.

For BodhAI, I worked with FastAPI, LangChain, Gemini, and a MongoDB vector store to improve semantic search latency in financial queries.

Multimodal only counts when ingestion is safe.

Voice transcription, document ingestion, session memory, and access-controlled indexing taught me that AI features need boundaries as much as models.

Agents need state, not just prompts.

I care about persistent run state, version history, scheduled actions, evidence trails, and live queries because agentic workflows need memory you can inspect.

Improvement should be visible.

TrustLoop tracks score deltas, repair cycles, execution modes, and best-version promotion so users can see how code gets better.

Edge AI, not just cloud demos.

I have worked on lightweight sleep-apnea event detection ideas using wearable-feasible signals, classical ML, and intervention-aware constraints.

Full-stack when needed, systems when it matters.

SwitchStream pushed me through real-time streaming, RTMP/WHIP, LiveKit, auth, Prisma state, webhooks, and chat moderation.

Realtime UX is an infrastructure problem.

Slow mode, block/unblock, participant kicking, and live status sync are product features until concurrency makes them systems problems.

I like software that crosses boundaries.

HALO connected Electron desktop capture, browser UX, Socket.io coordination, AWS uploads, and AI video intelligence in one workflow.

AI should remove creator chores.

HALO used transcription, summaries, generated titles, and descriptions to turn raw video capture into publishable context faster.

Pandas is still a serious systems tool.

Trade-to-holding reconciliation taught me to map transactions onto last known positions before computing exposure deltas and PnL attribution.

Deterministic keys beat clever matching.

For enterprise data workflows, I prefer stable identifiers, repeatable merges, and inspectable validation over fuzzy magic that cannot be audited.

Hardware makes software honest.

DRDO embedded work forced me to think about thresholds, ADC mapping, state machines, latency, and safety-critical behavior under real constraints.

Applied ML is mostly constraints.

The useful question is not whether a model works once, but whether the signals, latency, failure modes, and validation path can survive deployment.

Good AI products are backend-heavy.

The impressive part is often queues, state, permissions, retries, evidence, and observability hiding behind a calm interface.

Currently studying the hard parts.

Agentic workflows, eval frameworks, retrieval systems, model reliability, production observability, and the economics of AI latency.

I like tools with sharp edges removed.

The best engineering interfaces make complex state scan fast: stage feeds, score deltas, failure evidence, and clear next actions.

Early-career, not unserious.

I am not borrowing credibility from fake praise. I am showing the systems I have built, the failures I look for, and the habits I am developing.

Not testimonials. Signals.