Code where the rules are clear. Agents where they aren't.
Data platforms and AI agents, built inside your team and running in your own cloud.
- Feasibility first
- Code by default
- Agents for judgment
- Built in your cloud
Works with the stack you already runClouds, data platforms, models and business apps
Eight systems, eight architectures
Each one has its own animated architecture, a working demo and a 3D component view. The featured one is where most engagements start.
RAG and agentic RAG
Answers from your documents with citations, access control and evaluation. OKF bundles for curated knowledge.
Data engineering automation
Ingestion, Silver mappings, validation and SAM creation, assisted by an LLM hosted inside your network.
Workflow automation and AI agents
Manual processes mapped, automated in code, and handed to agents only where judgment is needed.
MCP servers and skills
Your skills, data and website served to any MCP-compatible AI tool from one governed endpoint.
Data warehousing
Star-schema warehouses, tuned for the reports your teams run every day.
Data lake and lakehouse
Medallion lakehouses in open formats, with quality gates, history and governed sharing.
AI harness and fine-tuning
Instructions, tools, guardrails and evaluations first. Fine-tuning when the evaluations call for it.
Analytical dashboards
Metrics defined once, reconciled against source, and shown in the BI tool you already use.
Problems we solve
Start from the problem, not the technology. Pick one to see what goes wrong today, how the work splits between code, an LLM and your people, and what changes once it runs.
Codesame input, same resultLLMreads, sorts and draftsPersonapproves what matters
Invoices stuck on hold
Invoices that don't match their purchase order land in a shared queue. Someone opens the PDF, the PO and the receipt side by side, emails the buyer and waits, while the vendor calls to ask where the payment is.
- CodeMatch in codeInvoice, PO and goods receipt matched with your tolerances.
- LLMExplain the gapAn LLM reads the invoice and the thread and writes why it doesn't match.
- CodeRoute by rulesEach exception goes to the buyer or approver who owns it.
- PersonApproveThe owner approves, rejects or asks the vendor, in Teams or email.
What changes
- Clean invoices post without anyone touching them
- Every exception arrives with a written reason and its documents
- Each decision is logged with who made it and why
The support inbox never empties
Every email is read by a person just to decide who should read it. Urgent requests wait behind routine ones, and the answer a customer gets depends on who picks it up.
- CodeCollectEmail, web forms and tickets pulled into one queue.
- LLMClassify and extractAn LLM tags intent and urgency and pulls out order and account numbers.
- CodeRoute and look upRules route by team and SLA; code fetches the order or account.
- PersonReplyAn agent sends the drafted reply, grounded in your help articles, or rewrites it.
What changes
- Urgent requests reach the right team first
- Drafts cite the policy they rely on
- Routing rules live in code your team can read and change
Answers are buried in documents
Staff search shared drives, PDFs and wikis for the current rule, or ask the one colleague who knows. Two people asking the same question get two different answers.
- CodeIndex with permissionsDocuments split and indexed, keeping each person's access rights.
- LLMSearch and answerAn agent searches, reads the passages and drafts an answer.
- CodeCheck citationsCode confirms every claim points to a passage the user may open.
- PersonFix the sourceDocument owners review flagged answers and correct the source.
What changes
- One answer, with the passage it came from
- People only see what they could already open
- Gaps in the documents show up as unanswered questions
Every new partner file is a project
Each new feed arrives with its own column names and codes. An engineer maps it by hand, writes the transformation and tests it, and the backlog grows with every partner.
- CodeProfile the fileEvery column read and profiled: types, formats, null rates.
- LLMPropose mappingsA private LLM maps columns to your standard, with a confidence for each.
- PersonSteward reviewA data steward approves the uncertain mappings.
- CodeGenerate and testPipeline code generated from the approved config and quality-checked.
What changes
- Mappings kept as reviewed config, not scattered code
- Engineers review a proposal instead of starting from a blank file
- No data leaves your network
Month-end lives in spreadsheets
Numbers are copied from the ERP into workbooks, adjusted by hand and emailed around. When a figure changes, nobody can say which version is right or why it moved.
- CodeModel onceERP data loaded into a warehouse with one tested definition per metric.
- CodeReconcileAutomated tests tie every total back to the source system.
- LLMDraft commentaryAn LLM explains month-over-month movements using tested numbers only.
- PersonSign offThe controller reviews the pack and signs it off.
What changes
- Every report uses the same metric definitions
- Differences surface before the review meeting, not during it
- Commentary never quotes an untested number
The CRM is always out of date
Reps finish a call and move straight to the next one. Next steps, new contacts and deal changes stay in notebooks, so the forecast is built on stale records.
- CodeCollect callsTranscripts and notes pulled in after each meeting.
- LLMExtract updatesAn LLM proposes field changes, next steps and new contacts.
- CodeValidateEach change checked against your CRM field rules.
- PersonRep confirmsThe rep confirms in one click before anything is written.
What changes
- Records updated the same day as the call
- Nothing written to the CRM without the rep's say-so
- Field rules applied the same way every time
Forms arrive as scans and attachments
Applications, contracts and claims come in as PDFs and phone photos. Someone keys the fields into the system by hand, and mistakes surface weeks later.
- CodeCaptureAttachments collected and run through OCR.
- LLMExtract fieldsAn LLM fills your schema from the document, scoring each field.
- CodeValidateFormats, totals and reference data checked in code.
- PersonReview exceptionsA person checks only the fields that failed or scored low.
What changes
- People check exceptions instead of typing everything
- Every field traceable to the page it came from
- Errors caught on arrival, not at month-end
Your AI tools can't see your systems
People paste data into chat tools to get help, or give up because the assistant knows nothing about your products, customers or processes.
- CodeWrap approved toolsAn MCP server exposes approved queries, documents and actions.
- CodeEnforce accessEvery call checked against the user's permissions and logged.
- LLMAssistant calls themClaude, Copilot or any MCP-compatible tool uses them when needed.
- PersonApprove actionsAnything that writes or sends waits for a person.
What changes
- One governed endpoint instead of copy and paste
- The same skills work in every AI tool your team uses
- A full log of what was asked and what was returned
Featured: data engineering automation with a private LLM
Ingestion, Silver mappings, validation and SAM creation, automated in code. An LLM inside your network, with no public access, drafts the parts engineers used to write by hand, and people approve them.
Validation run Passed with warnings
- member_id is not null100%
- diagnosis codes are valid ICD-1099.98%
- NPI passes Luhn check100%
- service_start_date not in the future318 failed
- paid_amount within plan limits100%
Rule drafter proposed 2 new checks from this batch's profile. Waiting for engineer approval.
Model: private endpoint in your VPC. 11 prompts logged, 0 left the network.
How an engagement runs
Forward-deployed means we work inside your team and your environment, from the first conversation to handover.
Assess
We sit with the people doing the work and map the process as it really runs, including the workarounds.
Feasibility
For each step we decide between deterministic code, an AI agent or a person. We check data access, security, environment and running cost, and give you a written go or no-go.
Build
We build in your environment, alongside your team, starting with the smallest piece that proves the approach.
Run and hand over
Monitoring, evaluation, runbooks and training, so your team can own what we built. We stay on call for as long as you need.
Where this applies
Our deepest experience is US healthcare. The same systems carry to any business with data spread across tools and work done by hand.
Healthcare payers and providers
Deep experience: reporting for 80+ US healthcare clients- Claims and EDI X12 standardisation
- CMS interoperability and FHIR gap analysis
- HEDIS and care-gap reporting
- Prior authorization assistants
Financial services
- Invoice and payment exceptions
- Reconciliation agents
- Regulatory reporting pipelines
- Policy and procedure assistants
Retail and supply chain
- Supplier feed ingestion
- Inventory and demand dashboards
- Order exception handling
- Product content served over MCP
Operations and public sector
- Document intake and extraction
- Case research agents
- Private LLMs with no internet access
- Knowledge assistants over policy
Who does the work
YoursSherpa engineering is led by Ashish Adhikari, a senior data and AI engineer who has built healthcare data platforms for US payers since 2022.
- 2025 to nowSenior Software Engineer, Abacus InsightsDatabricks pipelines on AWS that standardise US healthcare data. Medallion architecture with Delta Lake, SCD handling, Glue and Airflow orchestration, Snowflake publishing.
- 2022 to 2025Senior Data Engineer, Cedar Gate TechnologiesAmazon Redshift warehouse holding 10M+ healthcare records a year, ETL for 500K+ claims a day, and validation frameworks.
- 2021 to 2022Remote AI and ML Specialist, CleveroAssessed where machine learning fits into application workflows.
- CertifiedDatabricks Certified Data Engineer Associate, 2024Spark, Delta Lake, Delta Live Tables, Workflows and Unity Catalog.
- Research and awardsPublished and recognisedResearch on GANs for anomaly and malware detection (2024), a paper at the 73rd International Astronautical Congress, Abacus Hackathon winner 2025, and Microsoft Imagine Cup 2021 world finalist.
Questions clients ask
Short answers. The feasibility call covers the rest.
What does forward-deployed engineering mean?
We work inside your team and your environment. We sit with the people doing the work, build in your cloud accounts under your security rules, and hand over code, runbooks and training at the end.
How do you decide between automation and an AI agent?
Step by step. If the rule can be written down and stays the same, it becomes code. If the step needs reading unstructured input, judgment or drafting, an agent does it through tools we define. If the action is costly or hard to undo, a person approves it first.
Can the language model run without internet access?
Yes. We host models on private endpoints in your cloud, such as Databricks Model Serving, Amazon Bedrock through PrivateLink, Azure OpenAI private endpoints or Vertex AI through Private Service Connect, or run open-weight models on your own hardware with vLLM. Sheet 02 shows this design.
Do you only work on Databricks?
No. Our deepest delivery experience is on Databricks, AWS and Snowflake, and every design on this site also runs on Azure, Google Cloud or a private cloud. The feasibility analysis picks the environment you already run.
What happens in the feasibility analysis?
We map the process or data, check access, security, hosting and running cost, and mark each step as code, agent or person. You get a short written report with the architecture, the risks and a go or no-go recommendation before anything is built.
Where do the code and data live?
Everything is built in your accounts and repositories, so the code, configuration, prompts, logs and data stay with you.
Can AI tools like Claude or ChatGPT use our website?
Yes, through an MCP server that exposes your content and approved actions as tools, with authentication and an audit log. Sheet 04 shows how.
Start with a feasibility call
Tell us about one process or data problem. In the first call we will say which parts we would automate with code, which need an agent, and which we would leave alone.
Send a short note through the contact form and we will set up the call.
What helps us prepare
- The process or system you have in mind, and who works on it today.
- Where the data lives: cloud, platform and main tools.
- Security or hosting rules we need to work within.