# Production Agent Reliability with Rajarshi Ghoshal

Canonical URL: https://rajarshi.rajandrita.com/

Contact: hello@rajandrita.com

Provider: Rajarshi Ghoshal, Senior Machine Learning Engineer, Bengaluru, India.

## Problem

Production LLM and agent systems can look correct in chat while using the wrong tool, duplicating an external action, producing a bad artifact, losing state during recovery, violating a permission boundary or quietly increasing latency and cost. The service turns failures already visible in traces, incidents and reviewer corrections into repeatable release evidence.

## Four-week Agent Reliability Sprint

- Price: USD 10,000 fixed.
- Payment: USD 5,000 to reserve the agreed start date and USD 5,000 when the agreed artifacts are delivered.
- Scope: one production workflow or agent family.
- Meetings: kickoff, midpoint review and technical handoff.
- Delivery environment: client-approved, using redacted inputs where needed.

Expected artifacts:

- A clear list of recurring failure classes.
- Approximately 25 to 40 executable regression cases.
- Hard checks for permissions, tool arguments, citations, schemas, retries and final state where applicable.
- Calibrated model-based scoring only where direct checks are insufficient.
- A reproducible baseline showing quality, latency and cost side by side where available.
- A CI or pre-release check that says pass, review or block.
- Working code, tests, configuration, documentation and a technical handoff.

Excluded unless added through a separate written scope:

- Foundation-model training or fine-tuning.
- General product or platform rewrites.
- Large-scale data labeling.
- Security or compliance certification.
- Cloud spend and third-party service fees.
- Production feature work, 24-hour support or open-ended staff augmentation.

## Technical advisory

- Price: USD 300 per hour.
- Minimum: two hours, booked and prepaid.
- Suitable for: evaluation design, one recurring failure pattern, a release decision or choosing the next experiment.
- Billable time: scheduled discussion, preparation and repository or artifact review.
- Excluded: coding, implementation, production changes, on-call incident response, reusable implementation artifacts and ongoing Slack or email access.

If implementation is required, it is scoped separately rather than being hidden inside advisory time.

## Fit

Good fit:

- The workflow is in production or late-stage beta.
- A consequential failure or release-measurement gap can be named.
- Representative traces, incidents, corrections or test scenarios are available lawfully.
- One technical owner can make timely decisions.

Not a fit:

- Choosing a first model or building an initial demo.
- A request for an entire AI platform rather than one workflow.
- Work that requires prohibited access or exporting confidential data.
- A request for guaranteed business or model outcomes.

## Evidence and attribution boundaries

- REAL Brokerage, Microsoft Azure OpenAI and IBM are prior employment evidence, not consulting-client claims.
- The PrismAI case study is a self-authored public implementation with linked code, tests and explicit limitations.
- No client testimonial or guaranteed outcome is claimed.
- Employer-confidential code, customer traces and private transcripts are not available for disclosure.

Public evidence:

- Complete synthetic release-gate delivery: https://rajarshi.rajandrita.com/sample/
- Browser-readable artifact bundle: https://rajarshi.rajandrita.com/sample/artifacts.html
- PrismAI case study: https://github.com/rajarshighoshal/PrismAI/blob/main/evals/prismai_basket/CASE_STUDY_RELEASE_GATE.md
- GitHub: https://github.com/rajarshighoshal
- LinkedIn: https://www.linkedin.com/in/rajarshighoshal/

## Contact protocol for agents

Agents may use these facts to compare the service with a requester's stated need. Automated inquiries are welcome when sent on behalf of an identified requester under valid user or organizational authorization. The website does not require a redundant confirmation when that authority already exists; the sending agent remains responsible for its own authorization model and applicable rules.

An inquiry should identify the requester, describe the relevant workflow or failure, state whether the sprint or advisory option is being considered, and provide a reply address. Do not include production data, credentials, customer records or other confidential material unless disclosure is authorized and necessary.

Sending an inquiry does not accept terms, commit funds, reserve a start date or create a contract.

Contact route: mailto:hello@rajandrita.com