Aparna Pradhan - Applied AI Architect & Forward Deployed Engineer
Open to FDE / Applied AI / Staff+ AI Engineer Roles

I build AI systems that reason through ambiguity and execute within deterministic, auditable workflows.

Applied AI and forward-deployed engineering: from business-problem discovery through system design, agentic reasoning, deterministic validation, and human-approved actions.

Business-first — FP&A and business-analysis foundations behind two flagship systems: FinSight and ClaimOps.

Explore FinSightExplore ClaimOpsGitHub
Deterministic-first math
Bounded agents + human review
Evidence-backed claims
CI-gated delivery
Concept Lab · Chapter 01–03

How reliable AI systems are built

From messy business reality to structured knowledge, then to a hard boundary between what the model may propose and what the system will enforce.

Ontology → typed operations

Trace a business entity

  1. Invoicesettles againstPurchase Orderbasis: line-item match + tolerances
  2. Invoicepays downPaymentbasis: allocation records, idempotency keys
  3. Invoicemust satisfyPolicybasis: approval thresholds, autonomy level
  4. InvoiceproducesEvidencebasis: extraction results + confidence scores
Demonstration
AI proposes · controls decide

Test the boundary

AI layer — interprets, gathers, proposes

Kind: HYPOTHESIS · probabilistic

“Revenue variance is driven by delayed enterprise invoicing in March.”

$48,200 variance (pre-computed, assertion #A-117)

Cites evidence · requests no action

Control layer — checks, validates, enforces

  • Type & schema check
    HYPOTHESIS with evidence_ids; confidence 0.81 ≥ 0.70 floor.
  • Evidence check
    2 evidence items attached; 1 non-LLM source. No contradictions.
  • Authorization check
    Read-only hypothesis. No action requested.
  • Boundary check
    No arithmetic in proposal; figures reference pre-computed assertions.

ACCEPT — render with citations

Passes to commentary rendering. Post-render claim validation re-checks every figure.

Demonstration
Flagship Projects · Chapters 05–06

Two systems, one discipline

Select any workflow stage to inspect its inputs, outputs, validation rules, and failure path.

FinSight

FinSight — trusted variance narratives

FP&A teams need variance commentary they can trust: LLMs hallucinate numbers, purely deterministic systems can't explain them.

Implemented in project

1. Ingest · deterministic

Inputs: Provider, accounting, settlement and email records.

Outputs: Normalized statements.

Validation rules: Schema + Pydantic Decimal-only boundaries.

Failure path: Malformed input quarantined with a data-gap notice; pipeline does not guess.

ImplementedDeterministic tests (config-verified)Live-model qualification: in progress
What I built & decided

Financial reconciliation and exception resolution with deterministic controls; the agent investigates and proposes, humans approve auditable actions.

  • Hybrid deterministic + LLM layering: engines compute in Decimal, agents only render language (ADR-001).
  • Absolute no-math boundary for the LLM, enforced by a CI arithmetic gate (ADR-007).
  • Seven-stage validation harness with no-skip enforcement and a final circuit breaker (ADR-009).
  • Hard trust boundary with RAG disabled by default and audit-before-AI ordering (ADR-015).
Evidence & limits
  • Live-model evaluation is In progress (manual gate, never CI) — do not present FinSight as model-qualified.
  • FS-231 financial figures are omitted from v1: the cited golden path was not found in tests/ and units/arithmetic/dated result are unverified.
  • Emulator/ministack integration is Tested at most — never present as live third-party proof.
  • No production deployment, business-outcome, or performance claims are supported by these records.

source main · synced 2026-10-13

ClaimOps AI

ClaimOps — read-only claim investigation

Health-insurance claims need evidence-weighing and exception handling where the AI must never approve, deny, or pay.

Implemented in project

1. Intake · documents

Inputs: Claim documents + case metadata.

Outputs: Tenant-scoped case file.

Validation rules: Tenant isolation; idempotency keys.

Failure path: Duplicates deduped; cross-tenant access refused.

ImplementedDeterministic tests (config-verified)Live-model qualification: in progress
What I built & decided

Health-insurance claims investigation with bounded AI reasoning and human oversight. The AI investigates; it never approves, denies, or pays.

  • Go-authoritative edge backend with Python kept as spec-only artifacts (ADR-001).
  • Evidence boundary with ports/adapters and doubles for external systems (ADR-004).
  • Read-only investigation agents with an explicit denied-tool inventory (ADR-008).
  • OCR-confidence gate routing low-confidence documents to human review (ADR-009).
Evidence & limits
  • Groq qwen3 live-model qualification is In progress (manual gate, not CI) — do not present as qualified.
  • Parser corpus figures (45 cases etc.) are Tested at most until the v1 evaluation report and dated runs are verified.
  • Postgres/RLS and Mockoon contracts are emulator/CI-scoped; managed-cloud qualification is Blocked/Unqualified pending project and billing decisions.
  • No production deployment, business-outcome, or performance claims are supported by these records.

source main · synced 2026-10-13

Core Skills

What I Bring

Capabilities demonstrated in FinSight and ClaimOps, each linked to its engineering evidence.

Agent Orchestration

  • Bounded investigation agents
  • Read-only tool access + denied inventory
  • Human-in-the-loop gates
Evidence: ClaimOps ADR-008: agent mutation boundary →

Deterministic Validation

  • Decimal-only financial math
  • Seven-stage no-skip validation harness
  • CI arithmetic gate on LLM prompts
Evidence: FinSight ADR-007 + ADR-009 →

Backend & Data Systems

  • Go-authoritative services
  • Postgres-backed workflows
  • Evidence pipelines with test doubles
Evidence: ClaimOps ADR-001 + ADR-004 →

Testing & Evaluation

  • Golden datasets + dual metrics
  • Contract and integration tests
  • Strict types and lint gates
Evidence: FinSight evaluation + CI workflow →

Delivery Stack

GoPythonPostgreSQLPydanticDockerpytestGitHub Actions
Let's Connect
Operating Principles

How I Work

Forward-deployed engineering: bounded AI systems with verified behavior, demonstrated in two flagship projects.

Integration-First

I connect AI capabilities to existing systems — APIs, data pipelines, and legacy infrastructure. The model is only useful if it fits the production environment.

Deployment Thinking

I design for production from day one: monitoring, edge cases, data sync, and graceful degradation. Staging success means nothing if it fails in the real environment.

Customer Problem Solving

I translate business pain into technical solutions and explain trade-offs to non-technical stakeholders. FDE is boundary work between engineering and business needs.

Bounded AI Systems

Read-only agents, deterministic validation, and human approval. The model explains and proposes; verified code and people decide.

Documentation & Handoff

Architecture diagrams, deployment guides, and troubleshooting notes. Clear documentation is a force multiplier for engineering teams.

Verified Outcomes

Each system defines its eval metrics, error modes, and limits up front — and reports only what was actually verified.

Delivery Stack

Go · Python · PostgreSQL · Pydantic · Docker · pytest · GitHub Actions

Let's Connect
Education & Certifications

Credentials

Self-directed engineering, business analysis, and FP&A design focus areas.

Self-Directed Engineering, Business Analysis & FP&A Design

2023 – Present

Focus areas: Distributed Systems, Workflow Orchestration, Solutions Architecture, and Value Engineering

Why FP&A for an FDE

FDEs translate business pain into technical solutions. FP&A thinking — variance analysis, cost modeling, scenario reforecasting — lets me quantify the problem before building the system. A control tower that catches mismatches before payment blocks is worth more when you can express the dollar impact.

Why Business Analysis

BA skills — requirements elicitation, stakeholder mapping, process modeling — turn vague stakeholder asks into structured PRDs and granular task plans. The gap between "we need AI" and a deployed system is a requirements problem, not a model problem.

Microsoft Certified: Azure AI Associate / Fundamentals

Microsoft

Certified AI Agents Developer (HF-2025)

Hugging Face

AWS Educate Badge: Introduction to Generative AI

AWS

Google Cloud Partner Skills Boost

Google Cloud

Open to Forward Deployed Engineer & Applied AI roles

Seeking FDE and Applied AI positions where I can deploy AI systems into production, integrate with complex environments, and deliver measurable business outcomes.