Senior Technical Engineer

SENIOR TECHNICAL ENGINEER

Reports to Mike Persichini, CEO
Location Remote · United States
Type [Full-time / contract] TBD
Compensation [Base] + [equity] + funded domain ramp
Start TBD

WHY THIS ROLE EXISTS

We need a hands-on senior engineer to own the build and ship daily across three services, Postgres and an LLM pipeline. The work is documented and sequenced; the autonomy and responsibility are substantial.

Much of the engineering is defensive: an entity firewall prevents guarantor data from entering the borrower’s spread; NOI source authority ranks operating statements above tax returns; and reconciliation surfaces variances. These controls are load-bearing. Understand them before simplifying them.

THE STACK

• kreditiq-agents: Python/FastAPI/Agno for extraction, scoring, compliance, memo generation, recovery and policy extraction. Primary focus.
• kreditiq-api: Go/sqlc system of record with 11 roles, 65 permissions and org-scoped queries.
• kreditiq-web: Next.js dashboard, pipeline, vault, settings and cited memo rendering.
• Data/AI: PostgreSQL 16 + pgvector; 44-table org-scoped schema; OCI Llama 4 Maverick; Cohere 1,024-dimension embeddings; schema-constrained validated calls.
• Platform: OKE, Istio, Keycloak OIDC and Kreuzberg; separate dev/prod clusters.
• Delivery: GitHub Actions gates on pytest, go test -race and vitest. No local cluster credentials; deployments and migrations use workflow dispatch.

WHAT YOU’LL BUILD

First: dev/prod prompt convergence. Prompts ship in the agents image. Prod carries two June 12 guidance hotfixes missing from dev; dev is about 27 commits ahead on citations, entity confirmation, firewall, recovery and JWT. Deploying dev as-is would regress guidance silently. Execute the agreed merge order, verify it and prevent recurrence.

You will then address these defined gaps:

1. Dead plumbing: agent_prompts writes UI→API→DB but Python never reads it; feature_flags has no consumers. Wire or remove them.
2. Regression breadth: add an operating-business golden package beside the current CRE package.
3. Shared defaults: decide how DSCR 1.25x, LTV 75%, FICO 680 and similar defaults behave without tenant config.
4. Prompt drift: generate settings from runtime text or label hardcoded copies as summaries.
5. Memo regeneration: decide whether document changes should prompt it.

Carry staged symmetric deletion, JWT verification, tenant auth and scrubbed prompts into prod.

LOAD-BEARING CONSTRAINTS

• Schema-constrained decoding—not temperature 0—is the reliability layer.
• OCI on-demand responses are capped at 4,000 tokens.
• Pydantic field descriptions are prompt surface and affect behavior.
• LLM-populated columns intentionally lack CHECK constraints. Raw responses persist in extraction_raw JSONB; normalized tables are an index, not source of truth.
• Truncate only LLM input; always store full extracted text.
• Explicit org filters and tenant context enforce tenancy, backed by isolation tests; RLS is secondary.
• The nine-agent UI taxonomy is presentation. Runtime is five version-controlled core prompts plus supporting calls.

REQUIRED SKILLS

• Production Python/FastAPI and LLM pipelines: constrained output, prompt versioning, evaluations and regression tests.
• Rapid productivity in Go, sqlc, RBAC and multi-tenant APIs.
• PostgreSQL/pgvector, live migrations, retrieval, cascades and isolation testing.
• Next.js ownership of dashboard, vault and cited memos.
• Kubernetes/CI/CD discipline; OKE/Istio preferred; testing across all three stacks.

Preferred: retrieval/embeddings; document AI/OCR; regulated systems; Keycloak/OIDC, JWT and tenant authorization.

Commercial lending knowledge is not required initially; we fund the ramp, but you must quickly judge extraction quality independently.

HOW YOU WORK

We value restraint, sound judgment under autonomy, production ownership, clear documentation, honest status reporting and curiosity about commercial credit. This is not a fit for someone who wants to rewrite the platform, trusts LLM output by default or needs a fully specified ticket before acting.

FIRST 90 DAYS

• Days 1–30: run all three stacks locally; read dev; explain the facts spine, analysis workflow and trust layer, including the absence of CHECK constraints.
• Days 31–60: complete prompt convergence through production, install a recurrence-prevention mechanism and deploy independent changes to dev.
• Days 61–90: resolve the dead plumbing with a written rationale; begin the second golden package; execute a production deploy and database migration independently.

EVALUATION AND PROCESS

Evaluation: Python/LLM depth; breadth across Go, Postgres, Next.js and Kubernetes; production judgment; engineering restraint; communication; references.

Process: intro call [30 min] → technical architecture conversation [60 min] → practical dev/prod divergence exercise [90 min] → code conversation about a consequential system [60 min] → final conversation with [name/role] and references.

TO APPLY

Send [CV/résumé and GitHub or equivalent] to [email] by [date]. Tell us about inherited code you wanted to delete but kept—and why.

ROLE TWO · SENIOR SALES ENGINEER

Reports to Mike Persichini, CEO
Location [Remote / hybrid — city] · [Travel, e.g., up to 25%]
Type [Full-time / contract] TBD
Compensation [Base] + [variable tied to qualified pipeline and closed-won technical wins] + [equity]
Start [Date] TBD

WHY THIS ROLE EXISTS

During the engineering transition, customer conversations must remain credible. Buyers are credit officers who ask whether DSCR is correct and where it came from; they need answers without routine engineering escalation.

Within two quarters, the role expands into onboarding, policy configuration, tenant scoring and the engineering relationship.

SELLING ACCURATELY

The accurate position is: tone and emphasis are configurable; formulas, structure and thresholds require versioned releases. State it clearly and still earn the business.

WHAT YOU’LL DO

Pre-sales and evaluation

• Discover credit policy, document mix, products, thresholds and LOS/core systems.
• Own demos, including live analysis of prospect or representative loan packages with testable citations.
• Lead proof-of-value engagements with defined success criteria and honest results.
• Complete security, vendor-risk and model-risk questionnaires.
• Represent the platform before skeptical credit and risk committees.

Configuration and onboarding

• Run credit policy upload and extraction; validate authoritative appraisal and cash-flow rules by product.
• Configure tenant scoring weights and thresholds, loan products and document requirements.
• Draft the 1,500-character Organization Guidance note for prompt style, emphasis and terminology.
• Clarify before contract signature where configuration ends and a code release begins.

Feedback loop

• Convert deal and evaluation failures into reproducible engineering input.
• Help source a second golden package reflecting real customer deal shapes.

REQUIRED SKILLS

• Commercial lending fluency: DSCR, LTV, NOI, global cash flow, debt yield, credit memos and core source documents. This is least substitutable.
• LLM literacy: constrained output, retrieval, nondeterminism, evaluations and hallucination controls.
• Multi-tenant SaaS architecture: org-scoped isolation, RBAC, OIDC/SSO and environment separation.
• SQL/Postgres fluency to investigate facts, provenance and deletion behavior.
• REST, auth, webhooks and ingestion literacy to scope LOS/core integrations.
• Reproducible issue reporting that isolates a failure to a document, page, fact and environment.

Preferred: bank/lender solutions experience; SOC 2, GLBA, FFIEC and SR 11-7 familiarity; code/platform literacy; document AI/OCR.

You do not need to have built an LLM pipeline or sold AI products. You must explain the system accurately; lender credibility matters more.

HOW YOU WORK

You are calm and specific with skeptical senior credit leaders; honest under commercial pressure; discovery-led; able to translate between underwriting and engineering; precise in writing; comfortable representing an evolving product without inventing certainty; and resourceful without an engineer on standby.

FIRST 90 DAYS

• Days 1–30: operate the full product; upload a package, run analysis, read the memo and trace any number to its source; explain the facts spine and entity firewall without notes.
• Days 31–60: own demos independently; complete a security/vendor questionnaire; document what is customer-configurable versus release-dependent.
• Days 61–90: run a proof of value with defined success criteria; configure one customer’s policy and scoring; deliver structured feedback to engineering, including a proposed second evaluation deal shape.

EVALUATION AND PROCESS

Evaluation: commercial lending credibility; LLM, tenancy and data fluency; customer-facing skill; accuracy under pressure; written communication; references.

Process: intro call [30 min] → credit memo/domain and technical conversation [60 min] → lender-panel demo exercise [60 min], including a scoring-formula configurability question → take-home vendor-risk response on AI model governance [~2 hrs.] → final conversation with [name/role] and references.

TO APPLY

Send [CV/résumé] and a short note to [email] by [date]. Tell us about a time you gave a prospect an accurate answer that made the sale harder—and what happened.