KDK Tax Compliance Engineering

• 01Added AY 2026–27 ITR-6 JSON import for company returns, covering schedules like business income, balance sheet, depreciation, capital gains, deductions, losses, MAT/MATC and advance tax.
• 02Fixed ITR-6 JSON round-trips that were losing data: representative country IDs, 194-IA buyer PAN/Aadhaar, firm type and negative intraday/F&O values.
• 03Fixed Schedule III fixed-asset imports that dropped repeated rows. In the QA sample only 3 of 21 rows were being saved; after the fix all 21 were. A follow-up merged matching opening and addition entries, which saved 41 assets from a 50-row sample.
• 04Smaller fixes in inventory notes, previous-year values, DTAA cess, foreign addresses, Zerodha ETF imports, audit report validation and Trial Balance import errors.

• 01IMS imports were failing on out-of-range amounts coming from the portal. One affected sync went from 0 saved rows to 466.
• 02Improved WhatsApp GSTR-2B report delivery, GSTR-1/3B report consistency, GST reconciliation and Tally sync.
• 03Lower-deduction certificate verification was checking every transaction row. Deduplicating it took one client from 4,519 checks to 9, and the query from 52.5 ms to 40.5 ms.
• 04Fixed challan-list pagination and TRACES fetching on the new portal. A client with 1,335 challans now loads in two pages.
• 05Connector fixes for e-pay challans, tax-audit filing sync, Zerodha imports and finding Chrome on macOS.

A customer-matching step went from 534 MB to 55 MB of memory, and database reads used about 35% less at production data size. I also batched GST, TDS and user reads and cut memory in Excel report generation, checking that the output sheets stayed the same.

The Income Tax, GST and TDS backends each read their prompts from per-workflow configs. I built an LLM Management page in the internal dashboard to browse and edit them. It checks the prompt and its required variables, saves to S3 and can reload the ECS service. I also added structured logs (workflow, config version, model, prompts) right before the model call in three AI workflows.

I also wrote a Claude Code /bugfix command that ended up in 12 repos. You give it a Zoho issue and it walks through tracing the code, reproducing the bug, finding the root cause, fixing and testing locally, and preparing the PR, using Zoho MCPs, Claude in Chrome and the GitHub CLI. It stops for approval before doing anything external, masks customer identifiers, never runs DML or DDL on remote databases, and sends architecture or business-rule questions to a person.