Applied AI systems engineering

Akhil
Pavuluri

I build systems that have to be correct under real conditions — Printish is the production proof.

Scroll to explore
Focus Areas

Reasoning · execution · monitoring

Deterministic reasoning

Structure decides, language only expresses intent

Decision graphs and semantic layers turn requests into explicit paths instead of soft guesses.

Execution systems

Execution is validated before anything moves

Typed plans, validators, and replayable runs replace model-led side effects.

Failure and drift intelligence

Failures stay visible before they spread

Causal attribution and drift monitoring make degradation inspectable while it is still actionable.

Philosophy

Good AI systems should be inspectable

The systems I trust are constrained on purpose: explicit contracts, visible failure paths, and enough structure to argue with.

I care less about model theater and more about whether the stack stays legible when production stops being polite.

Selected Work

Featured systems, pulled forward

01 / 03Causal inference · Execution graphs · Root cause
Causal AI System

TraceAI

Walks pipelines backward to find what actually caused a failure

Walks pipelines backward to find what broke. Initial ranking failed because undamped merge aggregation let the wrong node win — fixed with edge weighting and fan-in damping.

Causal graphsLineageFastAPIPostgreSQL

Why it stands out

Failure pressure

Incomplete instrumentation makes the graph sparse; attribution degrades to coarse buckets until edges exist.

Trade-off

Chose explainable graph math over deep learning attribution—better auditability, less expressive on unstructured text.

Open case study
02 / 03Shipped product · Non-AI · Full-stack
Production system (non-AI)

Printish

Print ordering platform — customer checkout, order tracking, and an admin dispatch view. Built and launched solo.

Customers upload artwork, pay, and track delivery in one clear flow. Admins manage dispatch, catch stuck jobs early, and keep updates honest when vendor or courier signals go stale. Designed, built, and launched solo.

TypeScriptNext.jsFirebaseRazorpay

Why it stands out

Failure pressure

Payment provider callbacks and retries can duplicate delivery—dedupe on provider ids and idempotent handlers.

Trade-off

Favored explicit states and boring CRUD over clever automation early—easier to debug under real traffic.

Open case study
03 / 03Deterministic analytics · Semantic layer · SQL
Deterministic Analytics

DecisionGraph

Schema-adaptive analytics without hallucinated SQL

Replaces brittle BI glue with a governed semantic graph: questions map to validated SQL, not model improvisation.

Semantic layerSQL generationGovernancePostgreSQL

Why it stands out

Failure pressure

Under-specified business terms map to the wrong metric ID—mitigated with explicit confirmation for ambiguous matches.

Trade-off

More upfront modeling work than a vanilla text-to-SQL toy; pays off in correctness and trust.

Open case study
Let's Connect

Open to roles where system reliability matters more than demo polish.

Available for remote and relocatable positions in systems and AI engineering.